Value at Risk for Linear and Non-Linear Derivatives ... at Risk for Linear and Non-Linear...
-
Upload
trinhkhanh -
Category
Documents
-
view
215 -
download
1
Transcript of Value at Risk for Linear and Non-Linear Derivatives ... at Risk for Linear and Non-Linear...
Value at Risk for Linear and Non-Linear Derivatives
by
Clemens U. Frei
(Under the direction of John T. Scruggs)
Abstract
This paper examines the question whether standard parametric Value at Riskapproaches, such as the delta-normal and the delta-gamma approach and theirassumptions are appropriate for derivative positions such as forward and optioncontracts. The delta-normal and delta-gamma approaches are both methods basedon a first-order or on a second-order Taylor series expansion. We will see that thedelta-normal method is reliable for linear derivatives although it can lead to signif-icant approximation errors in the case of non-linearity. The delta-gamma methodprovides a better approximation because this method includes second order-effects.The main problem with delta-gamma methods is the estimation of the quantile ofthe profit and loss distribution. Empiric results by other authors suggest the use ofa Cornish-Fisher expansion. The delta-gamma method when using a Cornish-Fisherexpansion provides an approximation which is close to the results calculated byMonte Carlo Simulation methods but is computationally more efficient.
Index words: Value at Risk, Variance-Covariance approach,Delta-Normal approach, Delta-Gamma approach, Market Risk,Derivatives, Taylor series expansion.
Value at Risk for Linear and Non-Linear Derivatives
by
Clemens U. Frei
Vordiplom, University of Bielefeld, Germany, 2000
A Thesis Submitted to the Graduate Faculty
of The University of Georgia in Partial Fulfillment
of the
Requirements for the Degree
Master of Arts
Athens, Georgia
2003
c© 2003
Clemens U. Frei
All Rights Reserved
Value at Risk for Linear and Non-Linear Derivatives
by
Clemens U. Frei
Approved:
Major Professor: John T. Scruggs
Committee: James S. Linck
Chris T. Stivers
Electronic Version Approved:
Maureen Grasso
Dean of the Graduate School
The University of Georgia
August 2003
Acknowledgments
I would like to thank my major professor, Dr. John T. Scruggs, for his patience,
guidance, and extraordinary support he provided during the process of writing this
thesis.
Furthermore, I would like to thank Dr. James S. Linck for his advice throughout
the whole year at the University of Georgia as well as for being on my committee. I
would also like to thank Dr. Chris T. Stivers for being on my committee.
I would also like to express my gratitude to the DAAD (German Academic
Exchange Service) for supporting my year at the University of Georgia through
a generous scholarship.
Mein besonderer Dank gilt meiner Familie, die immer fur mich da ist, meine
Entscheidungen unterstutzt und mir meine Ausbildung ermoglicht hat.
iv
Table of Contents
Page
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
Chapter
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 Definition of Risk . . . . . . . . . . . . . . . . . . . . . 4
2.2 The Risk Model . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Derivative Pricing Models . . . . . . . . . . . . . . . 8
2.4 The Greeks . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Synthetic Options . . . . . . . . . . . . . . . . . . . . . 19
2.6 Payoff Functions . . . . . . . . . . . . . . . . . . . . . 22
3 The Variance-Covariance Value at Risk Approach . . . . . 26
3.1 Definition of Value at Risk . . . . . . . . . . . . . . . 26
3.2 Implicit Value at Risk Assumptions . . . . . . . . . . 28
3.3 Value at Risk for Single and Multi Assets . . . . . 29
4 Parametric Linear Approximation Value at Risk Models . 35
4.1 The Delta-Normal Approach . . . . . . . . . . . . . . 35
4.2 Risk Factor Coverage . . . . . . . . . . . . . . . . . . 40
v
vi
4.3 Advantages and Shortcomings . . . . . . . . . . . . . 42
4.4 Examples of the Delta-Normal Approach . . . . . . 45
5 Parametric Non-Linear Approximation Value at Risk
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.1 Introduction to Second-Order Effects . . . . . . . 52
5.2 Quadratic Models . . . . . . . . . . . . . . . . . . . . . 54
5.3 The Cornish-Fisher Expansion . . . . . . . . . . . . . 61
5.4 Alternative Approaches . . . . . . . . . . . . . . . . . 62
5.5 Delta-Normal vs. Delta-Gamma Approaches . . . . . 65
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Glossary of Notation . . . . . . . . . . . . . . . . . . . . . . . . . 74
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
List of Figures
2.1 The Delta of a Call Option vs. the Asset Price and the Passage of Time 15
2.2 The Gamma of an Option vs. the Asset Price and the Passage of Time. 18
2.3 Linear and Non-Linear Payoff Functions. . . . . . . . . . . . . . . . . 24
4.1 “Delta-only” Approximation vs. Black-Scholes Value of a Long Call
Option Position. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.1 Comparison between the Delta-Normal and the Delta-Gamma
Approximation with the Black-Scholes Model. . . . . . . . . . . . . . 69
5.2 Absolute Approximation Error of the Delta-Normal vs. the Delta-
Gamma Approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
vii
List of Tables
2.1 Data for a Hypothetical Call Option . . . . . . . . . . . . . . . . . . 14
2.2 Hypothetical Call and Forward Data. . . . . . . . . . . . . . . . . . . 23
2.3 Relationship between Instrument and Underlying Price or Rate. . . . 25
5.1 Statistical Properties of an Option vs. its Underlying Asset . . . . . . 56
5.2 Statistical Properties of the Delta-Normal vs. the Delta-Gamma
Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
viii
Chapter 1
Introduction
Financial disasters with derivatives in the early 1990’s such as Barings, Metallge-
sellschaft, Orange County, and Daiwa have led to considerable concern among reg-
ulators, senior management, and financial analysts about the trading activities of
companies, financial institutions, and banks. The causes for those debacles, among
others, range from rogue traders for Barings and Daiwa to misunderstood market
risks for Metallgesellschaft and Orange County.
Jorion [12] observes increased volatility in financial markets since the early 1970’s.
Among many other factors the fact that more and more assets, such as bank loans,
became liquid and tradable due to securitization led to more volatile markets. There-
fore, controlling market risk became very important for financial and non-financial
institutions, as well as regulators.
It can be argued (see [12]) that the development of Value at Risk (VaR) is a
direct response to the disasters in the early 1990’s. Eventually, the VaR approach
became an industry-wide standard for measuring market risk. To determine how the
VaR approach is different from other risk measures, Dennis Weatherstone, former
chairman of J.P.Morgan, made the comment, “at close of each business day tell me
what the market risks are across all businesses and locations.” Saunders and Cornett
[19, p.235] describe the above comment in the following manner, “what he wants is a
single dollar number which tells him J.P. Morgan’s market exposure the next business
day especially if that day turns out to be a bad day.” VaR provides a consistent
1
2
and comparable measure of risk denominated in currency units (given a confidence-
level and time horizon) across all instruments, business desks, and business lines.
Managers think of risk in terms of dollars of loss. That makes a VaR measure more
effective, such as communicating the risk exposure to senior management.
Most VaR implementations consider linear exposure to market risks. As we will
see later, linear approximations can be inaccurate for non-linear positions such as
options. However, the rapid growth of derivative markets reveals that derivatives
are widely used for hedging and speculation and non-linear exposures to risks are
common characteristics in portfolios. Therefore, it is important to be aware of the
implication of non-linear risk exposure. This paper attempts to answer the question
of whether standard parametric VaR approaches, such as the delta-normal and the
delta-gamma approach are appropriate for derivative positions such as forwards and
options.
The outline of the paper is as follows: Chapter 2 presents the underlying risk
model which illustrates the set of assumptions about the risk factors as well as
derivative pricing models for forwards, futures, and European-style options. Chapter
3 provides an introduction to the VaR methodology. I will also discuss the implicit
assumptions of standard VaR models.
Chapter 4 discusses the delta-normal method which is based on a first-order
Taylor series expansion. I will conclude that the delta-normal method is reliable for
portfolios with linear risk exposures but can lead to significant approximation errors
for portfolios with non-linear risk exposures. Non-linear positions such as options
will be discussed in chapter 5. The delta-gamma method provides a better approx-
imation because it models second-order effects. Since the relationship between the
value of the position and normally distributed risk factors is non-linear, the distri-
bution of the position is non-normal. I will show that this distribution exhibits high
skewness and excess kurtosis which makes a VaR calculation using the delta-normal
3
method inaccurate. The main difficulty with delta-gamma methods is the estimation
of the quantile of the profit and loss distribution. Different methods for estimating
this quantile are presented including the Cornish-Fisher expansion. In addition, a
method which matches the known moments of the distribution to one of a family
of distributions known as Johnson distributions and using a central χ2-distribution
to match the first three moments is presented. Empirical results by Pichler and
Selitsch [17] suggest the use of a higher moment Cornish-Fisher expansion. Delta-
normal approaches are appropriate for linear derivatives and for positions that reveal
only weak curvature. The delta-gamma approach offers a significant improvement
over the delta-normal method in the case of non-linear derivatives. In particular, the
delta-gamma method, when using a Cornish-Fisher expansion, provides an approxi-
mation which is close to results calculated by Monte Carlo Simulation methods but
is computationally more efficient.
Chapter 2
Background
2.1 Definition of Risk
Jorion [12] defines risk as the volatility of unexpected outcomes. Of the many sources
of risk a financial institution is exposed to, I will focus on financial risk and, in
particular, market risk. According to J.P. Morgan [13, p.2], market risk can be
defined as a “risk related to the uncertainty of a financial institution’s earnings on
its trading portfolio caused by changes in market conditions such as the price of an
asset, interest rates, market volatility, and market liquidity.” Market risk can occur
in terms of interest rate, foreign exchange, and equity price risk. Those risks will be
explained in more detail in the following paragraph which is based on Crouhy et al.
[5, pp.178].
Interest rate risk is the risk that occurs due to the fact that changes in
market interest rates have effects on the value of a fixed-income security.
The equity price risk can be further split into two sources, systematic risk
and idiosyncratic risk. Systematic, or market risk, refers to an sensitivity
of an instrument or portfolio value to a change in the level of broad stock
market indices. Idiosyncratic, or specific risk refers to that portion of a
stock’s price volatility that is determined by characteristics specific to
that firm. These specific risks could be its line of business, quality of
management, or a breakdown in its production process. As a result of
4
5
portfolio theory, idiosyncratic risk can be diversified away, while market
risk cannot. Foreign exchange risk is one of the major risks faced by
large multinational companies. It is caused by imperfect correlations in
the movement of currency prices and fluctuations in international market
interest rates.
To briefly conclude, market risks arise from open, i.e. not hedged or imperfectly
hedged, positions.
2.2 The Risk Model
2.2.1 Distribution of Risk Factors
At first, I will define a set of assumptions for the risk factors. There will be additional
assumptions when discussing the different VaR models which will be explained in
chapter 3.
Assumption 1:
Continously compounded returns on risk factors or proportional changes in risk fac-
tors follow a multivariate normal distribution (which is also called a joint-normal
distribution).
The following line of argument is based on Crouhy et al. [5]. Prices are assumed
to be log-normally distributed, so that log-returns, Rt, during the period (t − 1, t)
can be defined as
Rt = ln
(St
St−1
)
︸ ︷︷ ︸geometric return
= ln
(1 +
St − St−1
St−1
)∼ ∆St
St−1︸ ︷︷ ︸arithmetic return
,
where St (St−1) is the spot price of a arbitrary asset at time t (t − 1), and ∆St is
the price change during the period, ∆St = St − St−1.
6
This equation contains definitions for both the arithmetic return and the geo-
metric return. According to Dowd [6, p.41], geometric returns are approximately
equal to arithmetic returns when dealing with returns over short horizons, such as
one day. However, the advantage of using geometric returns is twofold. First, they
are more economically meaningful than arithmetic returns. If geometric returns are
distributed normally (i.e. prices are log-normally distributed), then the price is non-
negative. The second advantage of using geometric returns is that they easily allow
extensions into multiple periods which means that returns can simply be added.
It is worth noting that assumption 1 can also be used in a slightly different way.
In general, prices or levels of risk factors are assumed to be log-normally distributed.
Therefore, log-returns, or in more general terms, continously compounded propor-
tional changes are normally distributed. Bonds or interest rates are one example of
how this assumption is used. It depends on whether the bond price or the under-
lying yield to maturity is assumed to be a risk factor. RiskMetrics suggests the use
of bond prices as risk factors, whereas delta-normal or delta-gamma methods usu-
ally consider the yield to maturity as a risk factor. This reveals the problem that
the price-yield relationship for a bond is convex and therefore needs to be approx-
imated. Therefore, considering bond prices as risk factors mitigates the problem of
non-linearity in the payoff function.
Assumption 2:
Returns between successive time periods are uncorrelated.
This is consistent with the efficient markets hypothesis. An efficient market is
a market in which prices fully reflect available information. If so, all price changes
must be due to news that, by definition, cannot be anticipated and therefore must
be uncorrelated over time. According to Jorion [12], we can therefore assume that
prices follow a random walk. In addition, it can reasonably assumed that returns
7
are identically distributed over time. This assumption is the basis for a convenient
property of a VaR measure, i.e. that VaR measures can be easily translated over
different time periods.
Critical Comment on the Assumptions:
There is empirical evidence (e.g. Brooks and Persand [3]) that many individual return
distributions are not normal but exhibit what is called fat tails or excess kurtosis.
That means that these distributions reveal a far higher incidence of large markets
movements than is predicted by a normal distribution. To be more precise, fat tails
should worry risk managers because they imply that extraordinary losses occur more
frequently than a normal distribution would lead them believe. To conclude, a normal
distribution is more likely to underestimate the risk of extreme returns, whereas
primarily extreme losses are of concern to risk managers.
However, when considering a large and diversified portfolio, there is a statistical
theorem that should lower concerns regarding to the assumption of risk factors being
normally distributed. The central limit theorem states that the independent random
variables of well-behaved distribution will possess a mean that converges, in large
samples, to a normal distribution. In practice, this implies that a risk manager
can assume that a portfolio has a normal return distribution, provided that the
portfolio is well diversified and the risk factor returns are sufficiently independent
from each other. Crouhy et al. [5, p.193] argue that there is empirical evidence
doubting whether returns are sufficiently independent.
The assumption of returns on assets following a multivariate normal distribution
is convenient. This refers to the fact that VaR measures can easily converted over
different time horizons and confidence levels.
Nevertheless, some empirical studies suggest the class of Student-t distributions
to be more adequate for modelling portfolio returns. Those distributions allow for fat
8
tails and are fully characterized by the mean µ, the variance σ of the portfolio return,
and by a parameter called the degree of freedom (or degree of leptokurtosis), υ. As
υ gets larger the Student-t distribution converges to the normal distribution. Jorion
[12] states υ to vary in the range 4 and 8. Assuming υ = 5 and a confidence level
of 99%, the VaR is 3.365 times standard deviation instead of 2.326 times standard
deviation if the distribution is assumed to be normal. In this case, the effects of fat
tails is clearly shown since extreme returns are more likely to happen.
2.3 Derivative Pricing Models
In the following two sections I will focus on pricing models for forward/futures
contracts, and for European-style options.
2.3.1 Pricing a Forward Contract
At first, consider a forward or a futures contract on an investment asset. Forward
contracts are easier to analyze than futures contracts because there is no daily set-
tlement. However, under certain assumptions it can be argued that the forward and
the futures price of an asset are very close to each other when the maturities for
the contracts are the same.1 According to Hull [10], a forward contract obligates the
holder to buy (long position) or sell (short position) an asset such as a stock for a
predetermined delivery price (the forward price) at a predetermined future date.
The value of a forward contract at the time it is entered is zero. Hull [10] argues
that the value of a forward contract at time t, Ft, can be derived from the no-arbitrage
1To be more precise, it can be shown that when the risk-free interest rate is constantand the same for all maturities, the forward price for a contract with a certain deliverydate is the same as the futures price for a contract with that delivery date. This argumentcan be extended to cover situations where the interest rate is a known function of time.
9
argument as follows:2
Ft = St − FT e−rτ , (2.1)
where St is the current price of the security or commodity to be delivered (i.e. spot
price), FT is the delivery price, and r is the continously compounded risk-free interest
rate, and τ = T − t as the time to maturity.
We can see that the price of a forward contract depends on the current spot price,
the delivery price, the risk-free interest rate, and the time to maturity. The time to
maturity and the forward price are certain which leaves the spot price of the asset
and the risk-free interest rate as risk factors. From a risk management perspective,
market risk managers are interested in changes of the value of the forward contract
when these risk factors change. I will discuss sensitivity measures which measure the
exposure of the value of derivatives to changes in their risk factors in section 2.4.
2.3.2 Pricing a European-style Option
The Black-Scholes option pricing model [1] tells us what the price of a European-
style option on a non-dividend paying stock should be if it is consistent with a
no-arbitrage equilibrium. This is a state that rules out profitable, riskless trades.
Hull [10] summarizes the underlying assumptions of the Black-Scholes model as
follows:
1. The stock price, St, follows a geometric Brownian Motion.
2. The short selling of securities with full use of proceeds is permitted.
3. There are no transaction costs or taxes. All securities are perfectly divisible.
4. There are no dividends during the life of the derivative.
2I will assume for simplicity that the asset underlying a forward contract provides noknown dividend yield. An example is a stock that pays no dividends.
10
5. There are no riskless arbitrage opportunities.
6. Security trading is continous.
7. The risk-free rate of interest, r, is constant and the same for all maturities.
According to the Black-Scholes Model, the value of an European option on a non-
dividend paying stock is dependent on five variables. These variables are the under-
lying stock price, St, the volatility of the underlying stock price, σ, the continously
compounded interest rate, r, the time to maturity, τ , and the strike price of the
option, X. Therefore, the value Vt of an option at time t is a function V (·) depen-
dent on five variables.
Vt = V (St, X, σ, r, τ) .
The Black-Scholes valuation equation for a European call option on a non-
dividend paying stock is then:
Ct = St N(d1)−X e−r τ N(d2), (2.2)
where
d1 =ln
(St
X
)+
(r + σ2
2
)τ
σ√
τand
d2 =ln
(St
X
)+
(r − σ2
2
)τ
σ√
τ= d1 − σ
√τ ,
and where N(x) is the cumulative distribution function (CDF) for a variable, x,
that is normally distributed with a mean of zero and a standard deviation of one.
The Black-Scholes equation tells us that the price of a call option should be equal
to that of a leveraged position in the underlying stock. This interpretation of the
Black-Scholes option pricing equation in particular will be explained in more detail
when discussing synthetic options later in this section.
11
2.4 The Greeks
As discussed in the previous section, both the price of a forward and the price of
an option contract depend on various risk factors. As we have seen, the value of a
forward contract depends on two risk factors whereas the price of an option contract
depend on three risk factors.3 Sensitivity measures of a derivative’s value to the
different dimensions of risk are commonly referred to as the “Greeks.” Most often
the Greeks are thought of only being the sensitivity measure of an option but some
of them can also be calculated for other kinds of derivative contracts.
These measures result from a comparative statics analysis. Comparative statics
analysis is concerned with measuring the sensitivity of a dependent variable of a
function on the exogenous variables. The Greeks are first- or second-order partial
derivatives of the value function with respect to each risk factor. They describe
how the value of an option changes when the underlying risk factor changes by one
unit leaving all else constant (ceteris paribus). For the sake of completeness, I will
describe all Greeks for an option in the following paragraphs, though for our purpose
I will mainly use delta and gamma. The parameters rho and vega are also important
for risk management considerations, whereas theta can be ignored when modelling
short time horizons.
The Greeks are “local” risk measures. The term local refers to the fact that they
measure the sensitivity of a derivative’s value to infinitesimal changes in market rates
around current rates. Wilson [22] remarks that these representations of the payoff
profile based in local measures may not be sufficient to fully characterize the payoff
function for large market events. This aspect will also be discussed in greater depth
in this section.
3As will be discussed later, it is questionable if the passage in time must be consideredas a “risk” factor. This is due to the fact that the passage in time is certain and theta-effectsare small.
12
2.4.1 Delta
The delta of a derivative, δ, is defined as the rate of change in the value of the
derivative with respect to the price of the underlying asset, St. In the case of a
forward or an option, the delta is the slope of the curve that relates the forward or
the option price to the underlying asset price.
The Delta of a Forward Contract
The delta of a forward contract is the first partial derivative of the pricing equation
(equation (2.1)) with respect to the price of the underlying asset. Therefore, the
delta of a forward contract is,4
δ =∂Ft
∂St
= 1.
We can see that the delta of a forward contract is always equal to one which is the
reason why a forward contract is referred to as a linear derivative. I will discuss the
difference between linear and non-linear payoff functions in greater depth in section
2.6.
The Delta of an Option Contract
The delta of an option, δ, is defined as the first partial derivative of the pricing
equation (equation (2.2)) with respect to the underlying asset price. The delta for a
long position in a call option and a long position in a put option can be expressed
as follows:
δCALL =∂V
∂S= N(d1) = N
ln
(St
X
)+
(r + σ2
2
)τ
σ√
τ
and (2.3)
δPUT =∂V
∂S= N(d1)− 1 = N
ln
(St
X
)+
(r + σ2
2
)τ
σ√
τ
− 1. (2.4)
4I will assume the situation that the asset underlying a forward contract provides noknown dividend yield.
13
The delta of a call option can have values between 0 and 1, whereas the delta of a
put can have values between −1 and 0. As can be seen in equation (2.3) for a call
option, delta depends on the “moneyness” of the option, represented by the fraction(
St
X
), the time to maturity, τ , the risk-free interest rate, r, and the volatility of the
underlying asset, σ. In contrast to the delta of a forward contract, the delta of an
option contract changes due to changes in these factors. Theoretically, delta remains
constant only for an instant.
An option’s delta is the foundation for delta hedging. The idea behind delta
hedging is that a position is constructed so that the deltas of the different securities
offset each other. A position with a delta of zero is referred to as being delta neutral.
As we have seen, delta changes due to parameter changes. Therefore, the hedge has
to be adjusted regularly. This is known as rebalancing or dynamic hedging.
Figure 2.1 shows how the parameter delta of a hypothetical European-style call
option changes when both the underlying asset price and time to maturity change.
The data for this hypothetical option position is summarized in Table 2.1. This
example illustrates multiple points including the delta of an option position is close
to zero when the option is out-of-the money, whereas the delta is close to one when
the option is deep in-the-money. The most important point to be considered is that
the delta changes with respect to fluctuations in the variables on which it is based
namely time to maturity and the underlying asset price. It can therefore be deduced,
whenever delta is used for the purpose of delta hedging or the delta-normal VaR
method it is crucial to keep in mind that delta must be updated frequently when
parameters change. These changes in delta are larger with the passage of time and,
in particular, when the option trades around its exercise price.
However, Figure 2.1 reveals another property of the parameter delta. It can
be seen that large changes in delta occur when the current price in the underlying
instrument is near to the exercise price. In other words, we should expect to see large
14
Table 2.1: Data for a Hypothetical Call Option
Parameter Parameter ValueStrike Price (X) $100Volatility of the asset (σ) 40%Risk-free interest rate (r) 2.5%Price range of the asset (St) $0 - $200Time Horizon (τ) 0 - 1 year
changes in delta for small changes in the underlying price when the option is at-the-
money. This is the definition of the Greek letter gamma, which will be explained in
the next section and describes the second-order effect in the relationship between
the value of the option and the asset price.
2.4.2 Gamma
If changes in the underlying asset price are small, the option’s delta provides a good
approximation for the change in the option’s value. However, when changes in St
are large, and in particular when gamma is large in absolute terms, delta is highly
sensitive to the price of the underlying asset. This is due to the fact that delta
is a linear approximation of a non-linear function for options. Therefore, the real
change in the option’s value might deviate significantly from the amount predicted
by a delta approximation. The gamma of a forward contract is always zero since
the payoff function is linear in the price of the underlying asset, which means that
∂2V∂V 2 = 0.
The gamma, Γ, of an option is the rate of change of the option’s delta with
respect to the price of the underlying asset. It is the second partial derivative of the
15
Figure 2.1: The Delta of a Call Option vs. the Asset Price and the Passage of Time
16
option’s value with respect to the asset price. Therefore, gamma measures curvature.
Γ =∂2V
∂S2=
N ′(d1)
St σ√
τ, (2.5)
with
N ′(d1) =1√2π
e−d2
12 ,
where N ′(·) is the probability density function (PDF) of a standard normal distribu-
tion. Note that gamma is the same for a call and a put option. The parameter gamma
of long positions always has a positive sign whereas gamma for short positions is
negative.
As can be seen in equation (2.5), gamma depends on the “moneyness” of the
option, represented by the fraction(
St
X
)in the term d1, the time to maturity, τ , the
market interest rate, r, and the volatility of the underlying asset, σ. As a result, the
gamma of an option contract is sensitive to changes in these factors. To illustrate
this I will use the same numerical example that was used when discussing delta.
The parameters for the hypothetical European call option on a non-dividend paying
stock are given in Table 2.1. Table 2.1 also presents the ranges over which the price
of the underlying asset and the time to maturity will change.
As we can see in Figure 2.2, gamma reveals a bell shape with respect to the
underlying asset price. This is due to the fact that gamma (see equation (2.5)) is a
standard normal probability density function of the parameter d1 scaled by the term
St σ√
τ .
An option tends to have a high gamma when it is trading at-the-money. For an at-
the-money option or a close to at-the-money option, gamma increases dramatically
as the time to maturity decreases. The figure shows that gamma changes abruptly
when the option is near to being an at-the-money option and the time to expiry
is close to zero. Short-life at-the-money options have a very high gamma, meaning
17
that the value of the option holder’s position is highly sensitive to jumps in the stock
price.
I conclude now that if changes in the underlying asset price are small, the option’s
delta provides a good approximation for the change in the option’s value. However,
gamma effects must be considered when changes in St are potentially large, and in
particular when the option trades close to its strike price or at-the-money. Otherwise,
the approximation error can be significant. This is an argument against delta hedging
for assets with non-linear risk exposure. Due to the fact that delta hedging focuses
only on first-order effects, second-order effects are left out. Therefore, a delta-neutral
position is not insensitive to large price changes in the underlying asset.
2.4.3 Theta
The theta of an option, Θ, is the rate of change in the option’s value with respect
to the passage of time. Theta is sometimes referred to as the time decay of the
portfolio. Theta is not considered a risk factor since there is no uncertainty about
the passage of time. However, theta depends on other risk factors. When the time
horizon is short, such as one day, many authors (e.g. Hull [10]) argue that theta can
be ignored because changes in the option’s value due to a theta effect are small.
ΘCALL =∂V
∂τ= −St N ′(d1) σ
2√
τ− r X e−rτ N(d2). (2.6)
ΘPUT =∂V
∂τ= −St N ′(d1) σ
2√
τ+ r X e−rτ N(d2).
2.4.4 Vega
The Black-Scholes option pricing model implicitly assumes that the volatility of the
underlying asset is constant. In practice, volatilities change over time. That means
that the value of an option is liable to change because of the movement in volatility.
18
Figure 2.2: The Gamma of an Option vs. the Asset Price and the Passage of Time.
19
The vega of a single European-style put or call option is the rate of change if the
value of the single option or portfolio changes with respect to the volatility of the
underlying asset.
ϑ =∂V
∂σ= St
√τ N ′(d1), (2.7)
where N ′(d1) is the probability density function (PDF) for a standard normal vari-
able (see equation (2.6)).
2.4.5 Rho
The rho of an option is the rate of change in its value with respect to the risk-free
interest rate, r:
ρCALL =∂V
∂r= X τ e−rτ N(d2). (2.8)
ρPUT =∂V
∂r= −X τ e−rτ N(−d2). (2.9)
As we have seen in equation (2.1), the price of a forward contract is also sensitive
to changes in the market interest rate. The rho for a forward contract is:
ρ =∂Ft
∂r= −τ FT e−rτ .
Therefore, he parameter rho, ρ, describes the sensitivity of the forward price with
respect to changes in the risk-free rate for this maturity.
2.5 Synthetic Options
2.5.1 Creation of Synthetic Options
Based on the Black-Scholes option pricing model, European-style options can be
created synthetically. The components of such a synthetic option are called building
20
blocks and the whole process of replicating an option with buildings blocks, according
to a certain model, is called reverse-engineering.
As mentioned, when dealing with a standard European-style call option on a non-
dividend paying stock, the Black-Scholes model can be used to create an equivalent
synthetic option. Recall, the Black-Scholes formula:
Ct = St N(d1)︸ ︷︷ ︸δ
− X e−rτ N(d2)︸ ︷︷ ︸short position in discount−bond
The synthetic option position can be constructed by buying δ shares of the under-
lying stock at price St. The stock purchase is financed by borrowing (X e−rτ N(d2))
dollars. Alternatively, this could also be seen as shorting a discount-bond with a
present value of (X e−rτ N(d2)) and a face value (X N(d2)). This bond matures in
T and hence, has a duration of τ .
Once, this is done, one can then calculate a VaR figure for the equity and bond
positions separately, and then the combination of the two gives us our mapped option
position. The only novel point to watch here is that the components of the synthetic
option will change with changes in the underlying price - and indeed, other factors as
well - so the mapping itself needs to be regularly updated to reflect current market
conditions. Mapping options is therefore a dynamic process. As we have seen when
discussing the Greeks of an option, an option’s delta changes with respect to changes
in the price of the underlying, the passage of time, and interest rates. Changes in
delta can differ severely in magnitude depending on the “moneyness” of the option,
i.e. with respect to the price of the underlying asset.
Hull [10, Ch.13] remarks that the position in the stock and the option is risk-less
for only a very short period of time. Theoretically, it remains only risk-less for an
instant. To remain risk-less it must be adjusted or rebalanced frequently.
21
2.5.2 Portfolio Insurance
According to Hull [10], portfolio insurance is defined as entering into trades to ensure
that the value of a portfolio will not fall below a certain level. The value of a long
portfolio of stocks can be insured against dropping below a certain level by simul-
taneously investing in put options. An alternative to buying an option is creating it
synthetically. This strategy involves taking a time-varying position in the underlying
asset such that the delta of the position is equal to the delta of the required option.
Creating a synthetic put option implies that at any given time a proportion δPUT
(see equation (2.4)) of the stock has been sold and the proceeds invested in risk-less
assets.
As we have seen in the previous section, delta is not constant. This means that
the synthetic option position has to be rebalanced frequently. Hull [10] remarks
that creating a put option synthetically does not work well if the volatility of the
index changes rapidly or if the index exhibits large jumps. An example of this is
the stock market crash in 1987. On Monday, October 19, 1987, the Dow Jones
Industrial Average Index dropped by over 500 points and portfolio managers who
had chosen to insure their portfolio with synthetic put options found that they were
unable to sell either stocks or index futures fast enough to protect the position. I will
mention aspects of market liquidity again when discussing the implicit assumptions
of standard VaR models.
As we have seen, the creation of synthetic options is based on the option’s delta.
Such a decomposition of an option into a risk-less position in a discount-bond and a
short position in the underlying asset can also be used for modelling and calculating
a VaR figure. We will see this and the limitations when discussing the delta-normal
approach.
22
2.6 Payoff Functions
VaR is a measure for market risk. It requires an understanding of how a position’s
value changes when its underlying risk factors change. Whereas RiskMetrics [13,
p.123] suggests a classification of positions into three different categories, namely
simple linear positions, linear derivative positions, and non-linear positions, I will
here categorize positions only as linear or non-linear.
Linear positions include long or short positions in underlying assets such as stocks
or foreign currencies or long and short positions in linear derivative contracts such as
a forward contract on a stock or on a foreign currency. Non-linear positions include
an option on a stock or on a foreign currency.
Figure 2.3 illustrates these two types of positions and, in particular, how the
value of a position varies with price changes of the underlying asset. This figure is
based on a numerical example that contrasts the payoff function of a long call option
on an asset (a non-linear exposure) with the payoff function of a long position in
a forward contract (a linear exposure) on the same asset with the same time to
maturity. The specific data for both positions is summarized in Table 2.2.
The straight line in Figure 2.3 represents a linear relationship (i.e. δ = 1) between
the position’s price and underlying asset. This line represents the payoff function of
the forward contract. We can see that a change in value of such a position can be
expressed in terms of the delta, i.e. the slope, of the underlying security.
The other line represents the payoff function of the call option. This line clearly
represents a non-linear relationship between the position’s value and the underlying
security. The payoff line is curved such that the position’s value can change dramati-
cally as the value of the underlying asset increases. We can see that when measuring
the sensitivity of an option contract by its delta, this is only a good approximation
23
Table 2.2: Hypothetical Call and Forward Data.
Position Parameter Parameter ValueLong Call Option Current Asset Price (St) $100
Strike Price (X) $100Volatility of the Asset (σ) 40%Risk-free interest rate (r) 2.5%Time Horizon (τ) 45 days
Long Forward Contract Current Asset Price (St) $100Forward Price (FT ) $100Risk-free interest rate (r) 2.5%Time Horizon (τ) 45 days
for small changes. When discussing the Greeks of options, the curvature of the line
is quantified by, and referred to, as the parameter gamma.
Table 2.3 summarizes the points already made in this section and gives examples
for both categories of payoff functions. The table is partly based on RiskMetrics
Technical Document [13, p.124]. It is worth noting that there are two approaches
to modelling interest rate securities, such as bonds or interest rate swaps. As can
be seen in the table 2.3, one way to model interest rate securities is to consider the
price of the bond or the swap price as the underlying risk factor. In this case, the
relationship between the position’s value and the underlying price is clearly linear.
Another way is to consider a representative market interest rate as the underling risk
factor. In this case, the relationship between a bond price and its yield to maturity
is convex or clearly non-linear. For convenience, fixed-income securities are usually
modelled using representative bond prices, and not yields, as risk factors to avoid a
non-linear relationship which would be more difficult to model.
24
-20 -10 0 10 20Change in Underlying Asset Price (in $)
-20
-10
0
10
20
Cha
nge
in P
ositi
on V
alue
(in
$)
Change in the Value of the CallChange in the Value of the Forward
Figure 2.3: Linear and Non-Linear Payoff Functions.
25
Table 2.3: Relationship between Instrument and Underlying Price or Rate.
Type of Position Security Underlying Price/RateLinear Stock Stock Price
Commodity Commodity PriceForeign Exchange (FX) FX RateBond Bond PriceInterest Rate Swap Swap PriceFX Forward FX RateFloating Rate Note Money Market PriceForward Rate Agreement Money Market PriceCurrency Swap Swap Price/FX Rate
Non-Linear Stock Option Stock PriceBond Option Bond PriceFX Option FX Rate
Chapter 3
The Variance-Covariance Value at Risk Approach
3.1 Definition of Value at Risk
Jorion [12, p.22] defines VaR as “the worst expected loss over a given horizon under
normal market conditions at a given probability level known as confidence level.” In
more formal terms, VaR describes the quantile of the projected distribution of asset
or portfolio (dollar) returns over a target horizon. A target horizon could be a single
day or a 10 day period for the purpose of regulatory capital reporting. If c was the
selected confidence level, VaR corresponds to the 1− c lower-tail level.
According to Crouhy et al. [5, p.187], VaR offers a probability statement about
the potential change in the value of a single position or an entire portfolio resulting
from a change in market factors over a specific period of time. It does not state,
however, how much actual losses will exceed the VaR measure. Instead, it states
how likely it is that the VaR measure will be exceeded.
There are two interpretations of VaR:
1. VaR = Expected profit/loss – worst case loss at a given confidence level c
2. VaR’ = Worst case loss at a given confidence level c.1
To be more precise, the VaR over a horizon H and a given confidence level c can be
expressed as:
VaR(H; c) = −α1−c σ V = −α1−c σ(∆V ), (3.1)
1VaR’ is known as the “absolute Value at Risk”
26
27
where the VaR measure is either calculated with σ, which is the volatility of returns,
or with σ(∆V ), which is the volatility of dollar returns. The expression σ(∆V ) = σV
defines the relationship between both volatility definitions.
VaR′(H; c) = −(α1−c σ + µ) V = −α1−c σ(∆V ) + µ V, (3.2)
where the parameter α1−c denotes the 1 − c quantile of the assumed distribution
of returns, µ is the expected return, and V is the marked-to-market value of the
position.
Therefore, the α1−c quantile of the distribution fulfils the condition that the
probability of an outcome x which is less than or equal to α1−c is 1 − c. More
formally this can be expressed as:
Pr(x ≤ α1−c) = 1− c
⇔ N(α1−c) = 1− c = 1%,
where N(·) denotes a cumulative distribution function (CDF). Since we are inter-
ested in α1−c, we need the inverse of the CDF, N−1(·), which leads to the following
expression:
⇔ α1−c = N−1(1− c) = N−1(1%) = −2.326.
Therefore, the α1% quantile of a standard normal distribution is −2.326.
The relative VaR definition and the absolute VaR definition lead to the same
result when the expected profit or loss of a position, µ, is zero. This assumption is
usually made that over a short period of time, such as one day, the returns of risk
factors follow a multivariate normal distribution with mean zero and a volatility of
σ. As we will see later, these two VaR definitions might differ when dealing with
positions that reveal a non-linear relationship to changes in their underlying risk
factors. This might be the case even though returns in the underlying risk factors
are normally distributed with mean zero. An example for this is an option position.
28
What we can see so far is that the VaR measure is equal to the standard deviation
σ of a position or a portfolio, multiplied by the value of the position or portfolio
V , and by the −α quantile of the distribution. Therefore, the term σV provides
the standard deviation of the profit and loss distribution and the quantile is an
alternative means of quantifying the distribution.
In these general definitions the −α quantile is not related to a specific distri-
bution. This means that for computing the VaR measure of a specific position the
−α quantile can be derived from a (standard) normal distribution or a Student-t
distribution with a certain degree of freedom.
I conclude that the task of deriving a VaR figure for a single position or a complex
portfolio can be scaled down to the problem of deriving a probability distribution of
returns or deviations of the position or the portfolio.
3.2 Implicit Value at Risk Assumptions
In this section I will present some implicit assumptions of standard VaR models. VaR
models rely on market prices. Standard VaR models implicitly assume perfect and
frictionless capital markets. In particular, these models assume that markets remain
liquid at all times and therefore, assets can be liquidated at prevailing market prices.
This means that these models assume that market prices are achievable transac-
tion prices. The risk that the liquidation value of a asset differs significantly from
the current mark-to-market value is called liquidity risk. This risk depends on the
prevailing market conditions. If assets are traded in deep markets most positions
can be liquidated easily with little price impact. In the case of thin markets, such
as exotic over-the-counter (OTC) derivatives markets, any transaction can quickly
affect prices. In times of market distress liquidity can dry up quickly. This can lead to
situations in which positions cannot be sold at all or to dramatically deflated prices.
29
Jorion [12, p.339] concludes that this marking-to-market approach is adequate to
quantify and control risk but is questionable if VaR is supposed to represent the
worst loss over a liquidation period.
VaR models are based on volatility and correlation estimates of risk factors
in order to aggregate diverse risky positions. An implicit assumption of standard
VaR models is that volatilities and covariances are constant throughout the forecast
period. The problem is that volatility and correlation maybe highly unstable over
time. However, it should be noted that when volatilities and correlations are esti-
mated under normal market conditions, these estimates may be not reliable in times
of market distress. Brooks and Persand [3] found that correlations between mar-
kets increase when volatility is high. In conclusion, VaR estimates based on normal
market conditions may seriously underestimate the riskiness of a portfolio when the
volatilities and correlations of risk factors change.
3.3 Value at Risk for Single and Multi Assets
In section 2.6, I separated securities into two different categories, namely linear
and non-linear positions. The variance-covariance approach can be applied to linear
positions, such as stocks and foreign currencies. For a more detailed list of securities
which belong to this category refer to Table 2.3.
Implicitly, we have already seen the equation for calculating a VaR measure for
a single position. An example would be a stock whose value depends on one risk
factor, i.e. the stock price. Equation 3.1 is the formal expression for calculating a
VaR measure in a single asset case.
Since a financial institution’s trading portfolio usually consists of a huge number
of different assets including positions in equities, bonds, and foreign currencies, it is
therefore essential to know how to derive a VaR measure for an entire portfolio.
30
It is worth noting that the variance-covariance approach considers risk factors in
terms of asset prices rather than in terms of underlying market rates, such as bond
yields. This is one of the differences between the variance-covariance approach the
delta-normal method which will be presented in the next section. The delta-normal
method considers a portfolio of assets in terms of their underlying risk factors rather
than in terms of their assets. This requires that assets which depend on more than
one risk factor must be decomposed into their underlying risk factors.
However, one could model a single position as a portfolio of risk factors. An
example would be an U.S. investor holding a long position in a bond issued by a
foreign government and denominated in a foreign (non-dollar) currency. In this case,
the value of the position is exposed to both interest rate risk and foreign exchange
risk. In order to calculate the VaR of that position, correlation effects between the
two risk factors would have to be considered in addition to the volatilities of the two
risk factors. Here, I will consider a portfolio of risk factors in terms of its assets.
The variance-covariance approach is an extension of the portfolio model by
Markowitz [15]. As long as market factors are not perfectly positively correlated
assets, the risk of a portfolio of assets is not just the sum of the risks of every
asset in portfolio but reveals diversification benefits (see Crouhy et al. [5, p.99]).
The variance-covariance approach assumes that the distribution of changes in the
portfolio value is multivariate normally distributed (MVN). The multivariate normal
distribution is completely characterized by its first two moments, the mean vector
and the covariance matrix. These can be estimated from historical data. Another
assumption is that the value of a position has a linear relationship to the underlying
risk factor (i.e. stocks, foreign currencies, futures, or forward contracts).
Dowd [6, p.63] briefly summarizes the two assumptions of the variance-covariance
approach as follows:
31
1. Returns follow a multivariate normal distribution.2
2. There is a linear relationship among the value of a position or portfolio and
its underlying risk factors, i.e. their prices.
To illustrate this approach, I will assume a portfolio with current marked-to-
market value V . Changes in value of the portfolio, ∆V , depend on deviations in K
underlying risk factors, ∆fk (k = 1, . . . , K).3 As mentioned, one could also think
of a single position which is exposed to changes in several underlying risk factors. A
change in value, ∆V , can then be defined as
∆V =K∑
k=1
∆fk. (3.3)
We can now define ∆xk as the proportional change in risk factor k4 so that
∆xk =∆fk
fk
.
Then substitute the definition of ∆xk into equation (3.3) which leads to
∆V =K∑
k=1
fk∆fk
fk
=K∑
k=1
fk ∆xk
I will now derive the volatility of ∆V , σ(∆V ), under the assumption that risk factors
are log-normal distributed and therefore, log-changes follow a multivariate normal
distribution as follows:
σ(∆V ) =
√√√√√K∑
k=1
f 2k σ2(∆xk) +
K∑
j=1
K∑
h=1;j 6=h
fj fh cov(∆xj, ∆xh), (3.4)
2To be more precise, the assumption is made that prices are log-normally distributedand therefore log-returns follow a normal distribution.
3Usually, K refers to the number of risk factors. Here, the risk factor is considered tobe the price of the asset. Therefore, if we have N assets K = N .
4If the risk factor k is for instance a stock, ∆xk can be interpreted as the rate of returnprovided by the stock over a time horizon, H.
32
where σ(∆xk) represents the volatility of proportional changes in risk factor k and
cov(∆xj, ∆xh) is the covariance5 between changes in the risk factors j and h. Now,
replacing cov(∆xj, ∆xh) by
cov(∆xj, ∆xh) = ρ(xj ;xh) σ(∆xj) σ(∆xh),
where ρ(xj ;xh) is the correlation coefficient6 between the proportional changes in risk
factor j and h. This leads to the following expression:
σ(∆V ) =
√√√√√K∑
k=1
f 2k σ2(∆xk) +
K∑
j=1
K∑
h=1;j 6=h
fj fh ρ(∆xj ;∆xh) σ(∆xj) σ(∆xh). (3.5)
Equation 3.5 looks rather complex and can be made more concise by using matrix
notation. I define f as a K × 1 column vector containing the current value or level
of each risk factor k, fk, and Σ as a K ×K covariance matrix for the risk factors.
Therefore,
σ(∆V ) =√
f ′ Σ f , (3.6)
where
Σ =
σ2(∆x1) cov(∆x1, ∆x2) . . . cov(∆x1, ∆xK)
cov(∆x2, ∆x1) σ2(∆x2) . . . cov(∆x2, ∆xK)
......
. . ....
cov(∆xK , ∆x1) . . . . . . σ2(∆xK)
and f =
f1
f2
...
fK
.
Equation (3.6) is the volatility of changes in a portfolio’s value. We can calculate
the VaR measure using equation (3.1) which multiplies the volatility estimate just
5Covariance is a measure of of the degree to which returns on two risky assets move intandem. A positive covariance means the asset returns move together. A negative covari-ance means they vary inversely.
6The correlation coefficient is a scale-free measure of linear dependence and lies in arange between −1 to +1. When equal to +1 (−1), the two variables are said to be perfectlypositively (negatively) correlated. When 0, the variables are uncorrelated.
33
derived with the quantile of the distribution. Alternatively, we can rearrange equa-
tion (3.5). The following transformation is based on Jorion [12, p.152].
Before substituting the formal definitions of VaR, mentioned earlier, into equation
(3.5), they will be restated below and solved for the volatility σ(·):
σ(∆V ) =VaRp(H; c)
−α1−c
and
σ(∆xk) =VaRk(H; c)
−α1−c
.
As mentioned, substituting these rearranged VaR definitions into equation (3.5)
leads to
VaRp =
√√√√√K∑
k=1
VaR2k +
K∑
j=1
K∑
h=1;j 6=h
ρ(∆xj ;∆xh) VaRj VaRh. (3.7)
It is possible to write equation (3.7) in a more convenient form using matrix notation.
Assume that a portfolio consists of K different risk factors, then the VaR measure
for the entire portfolio VaRp(H, c) can be calculated as follows:
VaRp(H, c) =√
VaR′ ρ VaR (3.8)
where ρ is a symmetric K×K correlation matrix among the risk factors and VaR is
a K×1 column vector consisting of individual VaR measures, VaRk for each position
in risk factor k.7
ρ =
ρ(∆x1;∆x1) ρ(∆x1;∆x2) . . . ρ(∆x1;∆xK)
ρ(∆x2;∆x1) ρ(∆x2;∆x2) . . . ρ(∆x2;∆xK)
......
. . ....
ρ(∆xK ;∆x1) . . . . . . ρ(∆xK ;∆xK)
and VaR =
VaR1
VaR2
...
VaRK
7All VaR measures must be calculated based on the same time horizon and the sameconfidence level.
34
It is obvious that the portfolio VaR takes diversification benefits into account due
to the correlation coefficient ρ(∆xj ;∆xh). As mentioned, the correlation coefficient
is a scale-free measure of linear dependence which lies in a range between −1 to
+1. When equal to +1 (−1), the two variables are said to be perfectly positively
(negatively) correlated. When 0, the variables are uncorrelated.
The following three cases show how correlation affect the VaR measure of a
portfolio. For the sake of simplicity, assume that the portfolio consists of two different
assets which could be a portfolio of long positions in two stocks. For the sake of
simplicity, short sales are not considered at this point.
ρ(∆x1;∆x2) = +1 : Perfectly positively correlated returns yield to VaRp = VaR1 +
VaR2. This shows that the portfolio VaR is equal to the sum of the individual
VaR measures, meaning that there are no diversification benefits.
ρ(∆x1;∆x2) = 0 : Uncorrelated returns lead to VaRp = [VaR21 +VaR2
2]12 . It is apparent
that this VaR measure is less than the sum of the individual measures. This
shows that diversification benefits reduce the exposure to market risks and the
result reflects the fact that with assets that move independently, a portfolio is
less risky than either asset.
ρ(∆x1;∆x2) = −1 : Perfectly negatively correlated returns lead to VaRp = [VaR1 −VaR2]. It is obvious that this VaR measure has the smallest value among the
three constellations. The two assets move inversely, and in the case that the
two VaR measures have the same value, the VaR measure can be reduced to
zero. This proves that, in this theoretical framework, the market risk exposure
is zero.
Chapter 4
Parametric Linear Approximation Value at Risk Models
4.1 The Delta-Normal Approach
4.1.1 Introduction to Delta-Normal methods
As we saw, the basic variance-covariance approach assumes that position’s values
reveal a linear relationship to their underlying risk factors. This assumptions holds
for instance for stocks and positions in foreign currencies as well as for linear
derivatives such as futures and forward contracts. Therefore, the variance-covariance
approach is clearly appropriate for securities that exhibit such a relationship, and
when changes in their underlying risk factor follow a normal distribution. The
variance-covariance approach considers risk factors in terms of asset prices rather
than in terms of underlying market rates, such as bond yields. The delta-normal
method considers a portfolio of assets rather in terms of their risk factors than in
terms of their assets. This requires that assets which depend on more than one risk
factor must be split off into their underlying risk factors.1
I will now turn to positions belonging to the category of non-linear payoff func-
tions. Examples for securities that reveal such a non-linear relationship are options,
callable or convertible bonds, and mortgage-backed securities.
The idea and assumption behind the delta-normal approach is that any kind
of relationship between changes in the position’s value and the risk factors can be
1Example two on page 46 illustrates the decomposition of a forward contract on aforeign currency into its risk factors.
35
36
sufficiently approximated by a first-order derivative with respect to changes in the
risk factor. To be more precise, the approximation will be accomplished by using
a first-order Taylor series expansion based on first-order partial derivatives of the
value of the position with respect to its underlying risk factors. Such a first-order
partial derivative can be interpreted as a sensitivity measure of the value of the
position with respect to changes in the underlying risk factors, such as a delta of
a derivative as was illustrated in chapter 2. Changes in the value of a position can
then be described in terms of the underlying risk factors.
The delta-normal method can be seen as an extension of the basic variance-
covariance approach. Alternatively, the variance-covariance approach can be seen as
a special case of the delta-normal approach where the delta is equal to one (δ = 1)
and the risk factors are assets. Both approaches lead to exactly the same result
when considering a linear relationship, where the first-order derivative is constant.
As will be illustrated in this section, the delta-normal approximation is reliable for
linear payoff functions but can lead to significant approximation errors when dealing
with curved payoff functions. In the latter case, the idea is to replace the non-linear
relation between asset values and underlying rates and prices, i.e. the risk factors,
with a linear approximation based on the asset’s delta.
The basic ideas of the delta-normal approach can be expressed mathematically as
follows: If a financial instrument has an equilibrium value V which can be expressed
as a function of K risk factors then the change in value, ∆V , can be approximated
by a first-order Taylor series expansion of the pricing equation as follows:
∆V =K∑
k=1
∂V
∂fk
∆fk + εk(1) =K∑
k=1
δk ∆fk + εk(1). (4.1)
The first-order partial derivative of the value function V with respect to the risk
factor fk,∂V∂fk
, can be viewed as a sensitivity measure of the value of the financial
instrument to changes or shocks in underlying risk factors and is usually referred to
37
as δk. Depending on the payoff function, the first-order approximation error, εk(1)
is zero in the case of a linear payoff function or greater than zero in the case of a
non-linear payoff function.
4.1.2 Delta-Normal Value at Risk for one Risk Factor
The least complex case is a single position that depends only on one (K = 1) risk
factor. Hence, equation (4.1) becomes
∆V =∂V
∂f1
∆f1 + ε1(1) = δ1 ∆f1 + ε1(1), (4.2)
where, again, ∆V is a change in value of the position V , δ1 is the rate of change of
the position V with respect to the underlying risk factor f1, and ∆f1 is the absolute
change in the underlying risk factor. Further, I define ∆x1 as the proportional change
in the underlying risk factor in one day so that
∆x1 =∆f1
f1
.
Substituting ∆x1 in equation (4.2) and omitting, for the sake of simplicity, the first
order approximation error, ε1(1), leads to
∆V ≈ f1 δ1 ∆x1.
The volatility of ∆V is then
σ(∆V ) = σ (f1 δ1 ∆x1) = f1 δ1 σ(∆x1). (4.3)
Once the standard deviation is derived, the calculation of the VaR over a holding
period H and to a confidence level c is straightforward and can be obtained by
substituting equation (4.3) into the formal VaR definition presented in equation
(3.1) which leads to
VaR(H; c) = −α1−c f1 δ1 σ(∆x1).
38
Comparing this final result to the VaR expression I derived earlier for the simple
variance-covariance approach, we can see that those two definitions differ in one
parameter. This difference is the sensitivity measure, δk, which measures the contri-
bution of a risk factor to the overall risk of a position. If it is now assumed that this
approximation is equal to one which is the case when considering a single stock, a
forward, or a futures contract, the result of the delta-normal approach is the same
as the variance-covariance approach. It is now apparent that the variance-covariance
approach is a special case of the delta-normal approach.
Example one on page 45 illustrates this method for a long position in a European-
style call option on a non-dividend paying stock. In the next section, I will discuss
the case when a position depends on more than one risk factor.
4.1.3 Delta-Normal Value at Risk for K Risk Factors
After discussing the case of a position which depends on one (K = 1) risk factor,
I will now generalize the discussion to any type of financial instrument, for which
the equilibrium value, V , can be expressed as a function of K risk factors, fk, k =
1, . . . , K. Recall equation (4.1):
∆V =K∑
k=1
∂V
∂fk
∆fk + εk(1) =K∑
k=1
δk ∆fk + εk(1).
Defining ∆xk as the proportional change in risk factor k,
∆xk =∆fk
fk
,
and substituting ∆xk in equation (4.1) leads to
∆V =K∑
k=1
fk δk ∆xk (4.4)
As could be seen when deriving the volatility for the simple variance-covariance
approach in equation (3.5), the expression becomes quite complex. The following
transformations can be made more concise when using matrix notation.
39
I now define R as a K × 1 column vector of proportional changes in the risk
factors and δ as a K × 1 column vector containing in each row the product of the
current price or rate of risk factor k, fk, and the sensitivity measure with respect to
each risk factor k. More formally,
R =
∆x1
∆x2
...
∆xK
and δ =
f1 δ1
f2 δ2
...
fK δK
.
Then, equation (4.7) can be expressed in matrix notation as follows
∆V = R′ × δ, (4.5)
where ∆V is still a scalar.
Under the assumption that risk factors are multivariate log-normally distributed
and hence, log-returns follow a multivariate normal distribution, the volatility of
equation (4.5) can be calculated by
σ(∆V ) =√
δ′ Σ δ, (4.6)
where
Σ =
σ2(∆x1) cov(∆x1, ∆x2) . . . cov(∆x1, ∆xK)
cov(∆x2, ∆x1) σ2(∆x2) . . . cov(∆x2, ∆xK)
......
. . ....
cov(∆xK , ∆x1) . . . . . . σ2(∆xK)
The matrix Σ is a K ×K covariance matrix that consists of all the variance terms
σ2(∆xk) of each proportional change in risk factor k on the main diagonal and the
covariances among the changes of the risk factor in the off-diagonal cells of the
matrix.
40
As was illustrated when discussing the simple variance-covariance approach, the
calculation of a VaR is now straightforward. The volatility estimate, just derived,
needs to be multiplied by a quantile of the probability distribution according to the
chosen confidence level.
4.1.4 Portfolio Considerations
After having discussed the two scenarios of a single position which is dependent
on either one or more than one risk factor in the two previous sections, I will now
illustrate the case of a portfolio of assets.
Britten-Jones and Schaefer [2] present an approach of how to approximate and
aggregate changes of K risk factors over n assets. The following formula is consistent
with the notation used earlier in this paper but differs from the notation in Britten-
Jones and Schaefer [2]. Also, the equation itself differs slightly from the original
equation presented in their paper. Here, for the sake of simplicity, the sensitivity
measure theta is considered to be as one of the regular risk factor, k.
Equation 4.1 can then be adjusted for the multi-asset case as follows
∆V =n∑
i=1
wi
K∑
k=1
fk∂Vi
∂fk
∆xk, (4.7)
where wi describes the quantities of asset i held in the portfolio. As can be seen,
the first-order sensitivity measures can be aggregated by using a weighted average
method over the sensitivity measures with respect to the same asset i.
4.2 Risk Factor Coverage
It was illustrated in chapter 2 how derivatives such as forwards, futures, and option
contracts can be priced. These pricing models reveal the risk factors on which prices
of derivatives are dependent. Recall that according to the pricing models stated in
41
equations (4.11) and (2.2), the values of a forward contract and an option contract
can be defined as follows:
Ft = V (St, r) and
Vt = V (St, σ, r, τ) ,
where V (·) is a value function and the variables are the same as in chapter 2.
In the previous section the delta-normal approach was presented which provides
a framework for approximating and relating changes in risk factors to changes in the
price of the position.
Which risk factors, and therefore which Greeks should be considered when using
a delta-normal approach? I will now summarize some arguments in favor or opposed
to some possible candidate factors.
Obviously, delta has to be considered for both forward and option contracts. As
we saw in chapter 2, delta has to be adjusted frequently for option positions and the-
oretically remains constant only for an instant. Furthermore, delta provides a reliable
approximation only when changes in the underlying asset price are small. Wilson
[22] argues that delta as a risk measure is not sufficient to manage an options book.
This is due to the fact that many options books are run with an explicit strategy of
being delta-neutral at all times. Paradoxically, many of the most popular VaR tech-
niques (e.g. the delta-normal or RiskMetrics methods) recognize only directional
price or delta risk. This is the reason why J.P. Morgan does not recommend that
their standard RiskMetrics technique be applied to portfolios that include options.
When considering a short period of time, the parameter theta is often left out.
Hull [10, p.354] argues that theta risk (which measures the expected change in value
due to the passage of time) can assumed to be zero. This assumption is useful, in
particular, when VaR is measured for a short period of time, such as a one-day time
horizon.
42
The Black-Scholes model assumes that the market interest rate and the volatility
of the underlying asset are constant. If this is true, no vega or rho effects have to
be considered. In practice, interest rates and volatility change. To mitigate these
limitations, Crouhy et al. [5, p.206] suggests modelling the change in value of an
option by including sensitivity measures with respect to changes in the volatility
of the underlying asset, σ, and with respect to changes in the interest rate, r, as
follows:
∆V =∂V
∂S∆S +
∂V
∂σ∆σ +
∂V
∂r∆r
∆V = δ ∆S + ϑ ∆σ + ρ ∆r.
This method treats implied volatilities as another risk factor, treats the vega as
another delta or directional price sensitivity, and incorporates it directly into stan-
dard methods such as RiskMetrics or the delta-normal model.
Nevertheless, there is a trade-off between the accuracy of the model and the
tractability of the model. If all risk factors were to be considered in a model, a large
number of volatility and correlation estimates would be needed.
4.3 Advantages and Shortcomings
Linear approximation models such as the delta-normal approach have a number of
attractions, such as that it keeps the linearity of the portfolio. This method provides
a tractable way of handling positions with non-linearity while retaining the benefits
of linear normality. The method of using a first-order approximation is plausible and
rather reliable in certain scenarios. According to Jorion [12, p.68], the delta-normal
approach is likely to be reliable when the portfolio is close to linear in the first place,
since only then can a linear approximation be expected to produce an accurate VaR
estimate.
43
According to Wilson [21, p.220], delta-normal methods are likely to be reliable
“if the time horizon is very short, e.g. intra-day, and if the products themselves have
relatively linear payoff profile [...] Thus, it may be very well suited for measuring
and controlling intra-day risks of a money market or foreign exchange book with few
option positions.”
Jorion [12] further concludes that linear approximations can seriously under-
estimate the VaR for options because they ignore second-order risk factors, such
as gamma risk with options and convexity with bonds. This shortcoming will be
described in more detail in the next chapter.
Wilson [22] notes that the approximation error arising from a delta representation
typically increases with the size of the market rate innovation (i.e. changes in risk
factors). Unfortunately, large market events or movements are exactly the kind of
scenario that risk managers are concerned about when calculating VaR. This can
be illustrated graphically when considering a long position in a European-style call
option. Consider again the numerical example from chapter 2. The option’s strike
price, the risk-free interest rate, and the asset’s volatility are given in Table 2.1.
In addition, the underlying asset trades at $100 and the call option has two days
to maturity. Figure 4.1 plots the first-order Taylor approximation of the option’s
value compared to the “true” option value based on the Black-Scholes option pricing
model for prices from $90 to $110 one day to the option’s maturity. The delta
approximation is based on the option’s delta two days before maturity. As we can
see, the approximation error increases with the size of changes in the asset price. To
achieve a closer approximation, I will consider second-order effects by including the
gamma of an option. This aspect will be discussed in next chapter.
44
90 95 100 105 110Underlying Asset Price (in $)
-5.5
-3.0
-0.5
2.0
4.5
7.0
9.5
Val
ue o
f the
Cal
l Opt
ion
(in $
)
Black-Scholes ValueDelta Approximation
Figure 4.1: “Delta-only” Approximation vs. Black-Scholes Value of a Long CallOption Position.
45
4.4 Examples of the Delta-Normal Approach
4.4.1 Example 1: Single Option Position
Hull [10] gives an example for a delta-normal model. One could think of an option
position consisting of one or several European-style options on the same asset, such
as a stock with current price S. This position can include long and short positions
on the same underlying asset.
As mentioned earlier, an option’s delta describes the rate of change of the value
of the option with respect to changes in the price of the underlying asset. Simul-
taneously, δ is defined as the rate of change of the value of the portfolio with the
underlying asset S.2 As mentioned when discussing the multi-asset case based on
Britten-Jones and Schaefer [2], the delta of the portfolio can be calculated from the
deltas of the individual options in the portfolio. Let wi denote the weight of option
i.3 The delta of the entire portfolio is given by
δ =n∑
i=1
wiδi, (4.8)
where δi is the delta of the ith option.
The change in the value of the option portfolio, ∆V , can now be approximated
as follows:
∆V = δ ∆S,
where δ is the rate of change of the value of the portfolio with S and ∆S is the change
in the underlying stock price. Further, I define ∆x as the proportional change in the
stock price in one day so that:
∆x =∆S
S.
2The notation and the derivation of this example is based on Hull [10].3To be more precise, the parameter wi describes a quantity of option contracts of option
i and the parameter δi is the delta of the ith option contract.
46
It follows that the an approximate relationship between ∆V and ∆x is
∆V = S δ ∆x. (4.9)
The standard deviation of ∆V is then
σ(∆V ) = S δ σ(∆x). (4.10)
This leads to a VaR figure of
VaR(H; c) = −α1−c S δ σ(∆x).
4.4.2 Example 2: Forward and Option Portfolio
The following example provides an illustration for a VaR calculation for a portfolio
consisting of a long position in a forward contract on a foreign currency and a
short position in a call option on the same currency. First, each single security
will be modelled before turning to the problem of calculating a VaR figure for the
entire portfolio of these securities. I will start with the forward contract on a foreign
currency. The notation used by Hull [10] will be used for the entire example.
Forward Contracts on Foreign Currencies
Forward and futures contracts are the simplest form of derivatives. According to
Table 2.3, their value is linear in the underlying spot rates and their risk can be
constructed from basic building blocks.
The underlying asset in a forward on a foreign currency is a certain number of
units of the foreign currency. Let St be the current spot price, measured in dollars,
of one unit of the foreign currency and Ft be the forward price, measured in dollars,
of one unit of the foreign currency. A foreign currency has the property that the
holder of the currency can earn interest at the risk-free rate prevailing in the foreign
country. I define rf as the value of a foreign risk-free interest rate for a maturity
47
τ with continous compounding and r as the domestic risk-free interest rate for the
same maturity.
According to Hull [10], the value of a forward foreign exchange contract, Ft, is
given by4
Ft = St e−rf τ − FT e−rτ , (4.11)
where FT is the contracted forward price at maturity. We can now see that the price
of a forward on a foreign currency is sensitive to changes in the spot rate, St and
changes in the domestic and foreign currencies.
This allows us to approximate a change in value of the forward contract using a
first-order Taylor series expansion, such as
∆Ft =∂Ft
∂S∆S +
∂Ft
∂rf
∆rf +∂Ft
∂r∆r. (4.12)
Calculating the first-order partial derivatives with respect to the underlying risk
factors leads to:
∂Ft
∂S= e−rτ ,
∂Ft
∂rf
= −τ St e−rf τ , and
∂Ft
∂r= τ FT e−rτ .
Substituting these partial derivatives into equation (4.12) leads to
∆Ft = e−rτ ∆S − τ St e−rf τ ∆rf + τ FT e−rτ ∆r.
4The valuation equation for a forward contract on a foreign currency differs from thepricing equation initially used for pricing a forward contract. This is due to the fact thata foreign currency has the property that the holder of the currency can earn interest atthe risk-free interest rate prevailing in the foreign currency.
48
Here, the risk factors are the spot rate as well as the domestic and foreign interest
rates. Alternatively, I will replace the interest rates by the price of zero-bonds, such as
U.S. T-Bills, for the same maturity. The price of a domestic zero-bond, P , is defined
as P = e−rτ , whereas the price of a foreign zero-bond, Pf , is defined as Pf = e−rf τ . I
will further define price changes in these zero-bonds in terms of interest rate changes
as follows:
∆P =∂P
∂r∆r = −τ e−rτ ∆r or ∆r =
∆P
−τ e−rτ
and
∆Pf =∂P
∂rf
∆rf = −τ e−rf τ ∆rf or ∆rf =∆Pf
−τ e−rf τ.
Now substitute these two expressions into equation (4.13) and replace ∆S with S ∆SS
,
∆P with e−rτ ∆PP
, and ∆Pf with e−rf τ ∆PP
. The resulting expression is
∆Ft =[St e−rf τ
] ∆S
S+
[St e−rf τ
] ∆Pf
Pf
−[FT e−rτ
] ∆P
P. (4.13)
Note that the forward position in the foreign currency can be decomposed into three
cash flows which are
1. a long position in a foreign currency, [St e−rf τ ],
2. a long position in a zero-bond denominated in the foreign currency, [St e−rf τ ],
and
3. a short position in a domestic zero-bond worth [FT e−rτ ].
The VaR calculation for this single forward position is now straightforward and can
be achieved by applying equation (3.5) to derive a volatility estimate for the forward
position and then simply multiplying this volatility estimate by the quantile for the
chosen confidence level.
49
Call Options on Foreign Currencies
For the purpose of valuing a European option on a foreign currency the Black-Scholes
option pricing model can be used as introduced earlier when assuming that the spot
rate follows a geometric Brownian motion process similar to that for stocks. This
leads to the following valuation equation5
Ct = St e−rf τ N(d1)−X e−rτ N(d2), (4.14)
where
d1 =ln
(St
X
)+ (r − rf + σ2
2)τ
σ√
τ
and
d2 = d1 − σ√
τ .
The parameter St describes the current spot exchange rate at time t, r is the conti-
nously compounded domestic interest rate for the maturity τ , rf is the continously
compounded foreign interest rate, X is still the strike price of the option, and σ is
the volatility of the exchange rate.
Considering the same risk factors, chosen for modelling changes in value of a
forward contract, leads to the following approximation for changes in the value of
the option:
∆C =∂C
∂S∆S +
∂C
∂rf
∆rf +∂C
∂r∆r. (4.15)
Deriving the sensitivity measures from valuation equation (4.14) with respect to the
underlying spot exchange rate, S, the domestic risk-free interest rate, r, and the
5This pricing equation for an option on a foreign currency differs from the option pricingequation used in chapter 2. However, a foreign currency is analogous to a stock providinga known dividend yield. The owner of a foreign currency receives a “dividend yield” equalto the risk-free interest rate, rf , in the foreign currency.
50
foreign risk-free interest rate, rf leads to:
∂C
∂S= δ = e−rf τ N(d1),
∂C
∂rf
= ρf = −τ S e−rf τ N(d1), and
∂C
∂r= ρ = −τ X e−rτ N(d2).
The next step is to substitute these partial derivatives into equation (4.15) and analo-
gously to the transformations done for the forward contract, replacing the parameters
∆S, ∆rf , and ∆r leads to
∆C =[St N(d1) e−rf τ
] ∆St
St
+[St N(d1) e−rf τ
] ∆Pf
Pf
−[X N(d2) e−rτ
] ∆P
P.
This equation is quite similar to that for a foreign currency forward. The differences
are that any building block, according to one risk factor, is multiplied either by the
term N(d1) or N(d2). Both expressions are risk-neutral probabilities. For example,
N(d2) represents the probability that the option will be exercised in a risk-neutral
world, i.e. the option expires in-the-money. One property from the Black-Scholes
option pricing model is that when the spot price becomes very large relative to the
exercise price, a call option is almost certain to be exercised. Since the spot price,
St becomes very large, both d1 and d2 become very large, and N(d1) and N(d2) are
both close to one and the expression becomes similar to the expression for a forward
contract. I will now turn to the case of a portfolio that consists of a long position in
a forward contract on a foreign currency and a short position in a European option
on the same foreign currency.
4.4.3 Portfolio of a Forward Contract and a Single Option
In this section I will discuss the scenario of a portfolio consisting of a long position
in a forward contract on a foreign currency and a short position in a European-style
51
call option on the same foreign currency. I assume the two contracts are based on the
same quantity of underlying currency, and have the same exercise price (X = FT ).
Therefore, the change in value of such a portfolio, ∆V , can be modelled as follows
∆V = ∆F −∆C. (4.16)
Substituting the expressions for ∆F and ∆C derived in this section into equation
4.16 leads to
∆V =[St e−rf τ
] ∆S
S+
[St e−rf τ
] ∆Pf
Pf
−[FT e−rτ
] ∆P
P
−[St N(d1) e−rf τ
] ∆St
St
−[St N(d1) e−rf τ
] ∆Pf
Pf
+[X N(d2) e−rτ
] ∆P
P.
After rearranging this equation, the final expression for ∆V is
∆V =[St e−rf τ (N(d1)− 1)
] ∆S
S+
[St e−rf τ (1−N(d1))
] ∆Pf
Pf
−[X e−rτ (N(d2)− 1)
] ∆P
P
As previously mentioned, in the case of a deep in-the-money call option the terms
N(d1) and N(d2) are close to one and the entire position is nearly risk-free (i.e.
∆V = 0).
This conclusion is tempting, but dangerous. We know from the Black-Scholes
model that the price of a European-style call option depends on more than the risk
factors used above. According to the Black-Scholes model, such an approach neglects
other crucial risk factors, such as vega risk and gamma risk which in particular will
be discussed in greater depth later.
Chapter 5
Parametric Non-Linear Approximation Value at Risk Models
5.1 Introduction to Second-Order Effects
As we saw in the previous chapter, the delta-normal approach assumes that the non-
linear relationship between changes in the position’s value and the risk factors can
be sufficiently approximated by a first-order Taylor series expansion based on first-
order partial derivatives of the value of the position with respect to its underlying risk
factors. This is clearly inappropriate and leads to a large approximation error when
changes in the underlying risk factors are large and when the payoff function reveals
strong curvature. The curvature of the payoff function of an option is measured
by the parameter gamma. As we saw in chapter 2 and, in particular, in Figure
2.2, gamma tends to be high when the option trades at-the-money and the time
to maturity is short. A high gamma means that the value of the option position is
highly sensitive to changes in the stock price.
Standard VaR models assume that risk factors follow a multivariate normal dis-
tribution. Assuming a linear relationship between the value of the position and the
risk factors implies that the profit/loss distribution of the entire position is also
normally distributed. Even if changes in the underlying risk factors are normally
distributed, non-linear risk exposures leads to a non-normal probability distribution
for the changes in the position’s value. The distribution can be skewed, meaning that
the distribution is asymmetric around the mean. When gamma is positive, the prob-
ability distribution tends to be positively skewed, whereas when gamma is negative,
52
53
the probability distribution is negatively skewed. A positively skewed distribution
has a thinner left tail than the normal distribution whereas a negatively skewed
distribution has a thicker left tail than the normal distribution.
Since VaR is critically dependent on the left tail of a probability distribution,
it is crucial to know whether the distribution is skewed or not. Compared to a
normal distribution, a positively skewed distribution has a thinner left tail which
pushes the VaR percentile to the right. In this case, the VaR figure will be too
high. Overestimating the probability of extreme negative events is unsatisfactory,
but not that alarming. However, the case of a negatively skewed distribution when
normality is assumed should cause concern. Compared to a normal distribution, a
negatively skewed distribution has a thicker left tail which pushes the VaR percentile
to the left. Assuming normality leads to a VaR figure that is too low compared
to the real probability distribution and leads to an underestimation of risk. Such
underestimation of risks increases when the curvature, i.e. gamma, of an option
position is high.
When using an option pricing model, such as the Black-Scholes model, gamma
can be derived using equation (2.5). As already mentioned, a long position in call
or put options always has a positive gamma, whereas a short position in call or put
options has a negative gamma. The lack of symmetry implies that the discretion of
the bias of VaR depends on whether one is net long or short in options.
Jorion [12, p. 207] mentions an economic interpretation of the skewed distribution
regarding to options as follows:
“note that the option distribution (of a long call) has a long right tail,
due to the upside potential, whereas the downside is limited to the option
premium.”
54
5.2 Quadratic Models
The basic idea of the delta-gamma approach is that a second-order Taylor series
expansion provides a better approximation than a first-order Taylor series expansion
as was used for the delta-normal approach. A second-order Taylor series expansion
is based on first-order and second-order partial derivatives of the position’s value
function with respect to the underlying risk factors.
5.2.1 Delta-Gamma Value at Risk for one Risk Factor
Consider a portfolio exposed to a single risk factor. A change in the value of this
position, ∆V , can be approximated by using a second-order Taylor series expansion.
The first partial derivative of the value function V (·) with respect to a change in
the underlying risk factor k, ∂V∂fk
, is delta. Analogously, ∂2V∂f2
kis the second partial
derivative of the value function V with respect to changes in risk factor k. When
applying this approach to an option, this is an option’s gamma.1 The second-order
Taylor expansion series is
∆V =∂V
∂fk
∆fk +1
2
∂2V
∂f 2k
(∆fk)2 + εk(2), (5.1)
where the term ∆fk ([∆fk]2) describes the change (squared change) in the underlying
risk factor. Finally, εk(2) is defined as the “second-order” approximation error.
Equation (5.1) can be applied to a position of options. The option’s value is
dependent on one asset whose current price is St. We can make use of the Taylor
series expansion and approximate the change in the value of the position ∆V by
using the position’s delta and gamma as the first and the second partial derivatives
1The gamma of a portfolio which consists of different options on the same underlyingsecurities can be achieved by aggregating the gammas of the single options using a weightedaverage. This method is similar to the approach used to aggregate deltas described inequation (4.8). When a portfolio consists of options on different underlying risk factors,correlation effects between the risk factors must be considered.
55
of the value function with respect to the price of the underlying asset:
∆V ≈ δ ∆S +1
2Γ (∆S)2. (5.2)
Setting
∆x =∆S
S
reduces equation (5.2) to
∆V = S δ ∆x +1
2S2 Γ (∆x)2. (5.3)
The variable ∆x can be interpreted as a rate of return of the underlying asset. The
variable ∆V is non-normal. According to Pichler and Selitsch [17], in the case of
only one risk factor the distribution is a noncentral χ2-distribution.
The next step is to derive the first three moments of this distribution. Assuming
that ∆x is normal with a mean of zero and a volatility of σ the following terms can
be derived based on equation (5.3):
E(∆V ) =1
2S2Γσ2,
E[(∆V )2] = S2δ2σ2 +3
4S4Γ2σ4, and
E[(∆V )3] =9
2S4δ2Γσ4 +
15
8S6Γ3σ6.
Mean and Variance
According to Hull[10], the next step is to define µV and σV as the mean and standard
deviation of the option position so that
µV = E (∆V ) =1
2S2Γσ2 and
σ2V = E
[(∆V )2
]− [E(∆V )]2 , with [E(∆V )]2 =
1
4S4Γ2σ4
56
Table 5.1: Statistical Properties of an Option vs. its Underlying Asset
Moment Option (∆V ) Returns $ ReturnsReturn ∆V ∆x = ∆S
S∆S
Mean 12S2Γσ2 0 0
Variance S2δ2σ2 + 12S4Γ2σ4 σ(∆x)2 S2σ(∆x)2
Skewness 3S4δ2Γσ4 + S6Γ3σ6 0 0
which leads to a variance term of:
σ2V = S2δ2σ2 +
3
4S4Γ2σ4 − 1
4S4Γ2σ4 = S2δ2σ2 +
1
2S4Γ2σ4
︸ ︷︷ ︸adjustment
.
In contrast to the linear case presented when discussing the delta-normal approach
where the gamma term is assumed to be zero, both the expected value and the
variance reveal adjustment terms. Since the gamma term is non-zero, the expected
value of ∆V is now clearly non-zero in contrast to the delta-normal approach where
the expected value was zero. The variance is also adjusted by the term 12S4Γ2σ4.
Skewness
Skewness is a measure of asymmetry in a random variable’s probability distribution.
The skewness coefficient of the probability distribution of ∆V , ξV , is defined as:
ξV =E [(∆V − µV )3]
σ3V
=E [(∆V )3]− 3E [(∆V )2] µV + 2µ3
V
σ3V
,
whereas the skewness is defined as
E[(∆V − µV )3
]= E
[(∆V )3
]− 3E
[(∆V )2
]µV + 2µ3
V
= 3S4δ2Γσ4 + S6Γ3σ6.
57
Table 5.1 summarizes the first three moments of the option position with the
moments of the underlying asset. I will now briefly summarize the points made so
far in this section:
1. Although it is assumed that the return on the underlying asset is distributed
with a mean of zero, the change in the option’s value is non-zero unless gamma
is zero.
2. The sign of the option’s mean will be determined by the relative magnitude
and sign of gamma and whether one is long or short in the option.
3. The variance of the change in the option’s value differs from the variance
of the return on the underlying instrument by an adjustment term 12S4Γ2σ4.
If gamma is zero, the position’s variance is the same as in the delta-normal
approach.
4. Finally, whether the change in value of an option position is positively or
negatively skewed depends on whether the option position is a long or a short
position.
5.2.2 Delta-Gamma Value at Risk for K Risk Factors
In this section, I will extend the univariate case presented in the previous section
to the case of a portfolio consisting of K risk factors. At first, I will will present
the general case when individual instruments in the portfolio may be dependent on
more than one market variable. This would be the case if the portfolio contained
diff swaps2 and choosers.3 The following expressions use the same notation as was
2A diff swap is a fixed-floating or floating-floating interest rate swap. One of the floatingrates is a foreign interest rate, but it is applied to a notional amount in the domesticcurrency. Floating-floating diff swaps are a vehicle for directly betting on spreads betweendifferent currency’s interest rates.
3According to Hull [10], a chooser option is an option where the holder has the rightto choose whether it is a call or a put at some point during its life.
58
used in the univariate case discussed above. I will start with deriving a second-order
Taylor expansion series for the case of K risk factors as follows
∆V =K∑
k=1
∂V
∂fk
∆fk +K∑
k=1
K∑
j=1
1
2
∂2V
∂fk∂fj
∆fk ∆fj + εk(2). (5.4)
Analogously to the case presented in the previous section, I will define ∆xk as the
proportional change in risk factor k as follows
∆xk =∆fk
fk
and substituting ∆xk in equation (5.4) leads to
∆V =K∑
k=1
∂V
∂fk
fk ∆xk +K∑
k=1
K∑
j=1
1
2
∂2V
∂fk∂fj
fk fj ∆xk ∆xj + εk(2).
The previous equation looks rather complex and can be made more concise by using
matrix notation:
∆V = δ′R +1
2R′ΓR, (5.5)
where the vector R is still a K × 1 column vector consisting of the proportional
changes of each risk factor k, ∆xk, and δ as a K × 1 column vector containing
in each row the product of the current price or rate of risk factor k, fk, and the
sensitivity measure with respect to each risk factor k. Therefore,
R =
∆x1
∆x2
...
∆xK
and δ =
f1 δ1
f2 δ2
...
fK δK
.
The matrix Γ denotes the K ×K matrix of gamma terms
Γ =
Γ1,1 Γ1,2 . . . Γ1,K
Γ2,1 Γ2,2 . . . Γ2,K
......
. . ....
ΓK,1 . . . . . . ΓK,K
with Γk,j =∂2V
∂fk∂fj
fk fj.
59
More specifically, the gamma matrix is the second derivative, or Hessian,4 of the
portfolio’s of position’s value function. The gamma terms on the main diagonal
(k = j) are conventional gamma measures in which case gamma is defined as ∂V 2
∂f2k.
In this scenario, the gamma of an option position describes the change in δk for a
change in risk factor k.
The off-diagonal terms are cross-gamma terms (k 6= j). They are often ignored
if the individual prices are functions of only one market price or if the cross-product
effects are trivial. Wilson [22] remarks that this is a potentially dangerous assump-
tion. For correlation-dependent products such as diff swaps or choosers, the cross-
product terms can be significant and should not be ignored.5 Jorion [12] notes that
in the case of an option position when there is more than one risk factor considered,
such as the asset price and the implied volatility there exists a cross-effect. This is
due to the fact that the delta of an option also depends on the implied volatility.
As in the univariate case, the moments of the distribution of ∆V must be deter-
mined first. According to Pichler and Selitsch [17], we can derive the moments of
the profit and loss distribution of the position as follows.
Mean and Variance
µV = E[∆V ] =1
2tr [ΓΣ] and
σ2V = E
[(∆V )2
]− [E(∆V )]2 = δTΣδ +
1
2tr [ΓΣ]2
︸ ︷︷ ︸adjustment
.
4Here, the Hessian is multiplied with the current level of the risk factors k and j.5See also Rouvinez [18].
60
The trace6 of the matrix [ΓΣ] is the sum of the K eigenvalues of [ΓΣ] and the trace
of [ΓΣ]2 equals the sum of the squared eigenvalues of [ΓΣ]. As was discussed in
the univariate case and in contrast to the linear case presented when discussing the
delta-normal approach, the expected value is non-zero and variance also reveals an
adjustment term. The adjustment term for the variance is the trace of [ΓΣ]2.
Skewness, Kurtosis, and higher Moments
Let X be the standardized value of ∆V
X =∆V − E[∆V ]√
σ2V
which leads to higher moments of X with r ≥ 3 as follows:
E [Xr] =12
r! δTΣ [ΓΣ]r−2 δ + 12
(r − 1)! tr [ΓΣ]r
(σ2V )
r2
.
In particular, the skewness can be derived for r = 3 as
E[X3
]= ξ =
3 δTΣ [ΓΣ] δ + tr [ΓΣ]3
(σ2V )
32
and the kurtosis can be derived for r = 4 as
E[X4
]= η =
12 δTΣ [ΓΣ]2 δ + 3 tr [ΓΣ]4
(σ2V )2
.
Recall that skewness is a measure of asymmetry in a random variable’s probability
distribution. Kurtosis describes the degree of “peakedness” of a distribution relative
to a symmetric normal distribution. If the distribution has a higher peak than the
normal distribution the kurtosis is greater than three. The distribution is then said
to be leptokurtic. Flat-topped distributions which have a kurtosis of less than three
are referred to as platykurtic. The standard normal distribution is mesokurtic which
6According to Zangari [13], the trace of a matrix is defined as the sum of the diagonalelements of the matrix.
61
means that the distribution is neither peaked, nor flat-topped and reveals a kurtosis
of three.
Having derived the moments of the profit/loss distribution of the entire position,
I will now focus in the next sections on approaches to determine the specific quantile
of this distribution.
5.3 The Cornish-Fisher Expansion
The idea behind the method presented in the following paragraphs7 is a direct
approximation of the required quantile of the distribution of ∆V based on the
Cornish-Fisher expansion8 around the quantile of a normal distribution. The con-
fidence interval parameter must be adjusted for the skewness and other distortions
from normality created by the presence of the gamma factor. This adjustment can be
achieved by using an approximation formula known as the Cornish-Fisher expansion
which provides a relationship between the moments of a distribution and its per-
centiles. The Cornish-Fisher expansion is based on the statistical principle that one
distribution (e.g. a chi-squared) can always be described in terms of the parameters
of another (e.g. a normal). This adjustment factor needs estimates of the skewness
and the kurtosis. However, suggestions regarding to the number of higher moments
that need to be considered vary. Hull [10] suggests the use of the first three moments
of the distribution, namely mean, variance, and skewness, whereas the Zangari [23]
suggests also including kurtosis. Pichler and Selitsch [17] even suggests the use of the
first six moments of the distribution. According to Zangari [23, p.9], we can simply
apply the adjustment by using our new confidence parameter and then proceed as
if the distribution were normal.
7This approach is based on the work of Fallon [8], Pichler and Selitsch [17], and Hull[10] which recommend approximate solutions based on the Cornish-Fisher expansion.
8See Johnson and Kotz [11] for details.
62
Once the moments of the distribution are derived, the estimation of the adjusted
quantile does not differ between the univariate and the multivariate case. Using the
first three moments of ∆V , the Cornish-Fisher expansion estimates the q-percentile
of the distribution of ∆V , wq as
wq = αq +1
6(α2
q − 1)ξV .
However, using the first four moments of ∆V , the Cornish-Fisher expansion esti-
mates the q-percentile of the distribution of ∆V , wq as
wq = αq +1
6(α2
q − 1)ξV +1
24(α3
q − 3αq)ηV − 1
36(α3
q − 5αq)ξ2V .
The parameter αq is the q-percentile of the standard normal distribution. According
to Zangari [23, p.9], we can simply apply the adjusted confidence parameter for the
VaR calculations. Therefore, the VaR of this position can be calculated as follows:
VaR(H; c) = −wq σV and (5.6)
VaR′(H; c) = −wq σV + µV V. (5.7)
In contrast to the delta-normal approach, the definition of the absolute VaR, VaR′,
and relative VaR, VaR, differ, since the mean of the distribution is now non-zero.
5.4 Alternative Approaches
As mentioned, the profit/loss distribution of a non-linear position is non-normal.
How should one estimate the quantile of this new distribution? In the academic lit-
erature, several approaches have been discussed. I will briefly mention these different
approaches.
63
The Delta-Gamma Normal Approach
Dowd [6] discusses an approach called the delta-gamma normal approach. The
essence of this method is to regard the extra risk factor (∆xk)2 as equivalent to
another independently distributed normal variable and treat them in the same way
as the first factor, ∆xk. The idea behind this approach is to preserve the normal
linearity as was the case of the delta-normal approach. Hull [10] remarks that the
first two moments, namely the mean and the variance, can be fitted to a normal
distribution which is better than ignoring gamma altogether. The assumption that
∆V is normal is less than ideal. It is based on the assumption that the quadratic
term (∆x)2 is also normally distributed. This is wrong since a squared normally
distributed variable is then itself a non-central χ2 variable. An approach such as
this relies on the assumption that ∆V is normally distributed was discussed in aca-
demic literature but is, according to Dowd [6, p.73], “clearly inadequate, being both
logically incoherent and unreliable and should therefore should ruled out of court.”
A moment-fitting Approach
Zangari [13] suggests an approximation using Johnson curves. Having obtained the
first four moments of the portfolio’s profit/loss (or return) distribution, Zangari [24]
suggests finding a distribution that has the same moments but whose distribution
is known exactly. He suggests:
“that it is beneficial rather than deriving the portfolio returns exact
distribution - which is intractable anyway - we find the statistical char-
acteristics, or moments, of its distribution. [. . . ] These moments depend
only on the price of the option, the current market prices of the under-
lying securities, the option’s Greeks, and the RiskMetrics matrix. Having
obtained the first four moments of portfolios return distribution, we find
64
a distribution that not only has the same moments, but also has a form
that is exactly known.”
Zangari [24] actually suggests matching those moments to one of a family of dis-
tributions known as Johnson distributions.9 Matching moments means finding a
distribution that has the same mean, standard deviation, skewness and kurtosis as
the portfolio’s return distribution. As mentioned in the quote above, this hypothet-
ical distribution will have the same moments as the “true” profit/loss distribution,
but unlike the true distribution will also have a known form. The VaR estimate can
then be calculated from this known distribution.
Matching moments to a family of distributions requires that ∆V is approximated
by transformed standard normal variable Y = f−1(Z), where Z is a standard nor-
mally distributed variable. According to Pichler and Selitsch [17], the specific choice
of the transformation function f−1 depends on the ratio of the square root of the
skewness and the kurtosis of ∆V . Furthermore, the following Johnson family distri-
butions are sufficient to cover all possible combinations of the first four moments:
Z = a + b log(
Y − c
d
)(Lognormal),
Z = a + b sinh−1(
Y − c
d
)(Unbounded), and
Z = a + b log
(Y − c
c + d−Rp
)(Bounded),
where a, b, c, and d are parameters with the restriction (c < Z < c+d) whose values
are determined by ∆V ’s first four moments and f(·) is a monotonic function.
Zangari [24] mentions that the parameter values for a, b, c, and d are chosen
through a moment matching algorithm that has an analytical solution in the log-
normal case. In the unbounded and bounded cases he suggests to make use of the
9The name Johnson comes from the statistician Norman Johnson who described aprocess of matching a particular distribution to a given set of moments.
65
iterative algorithm by Hill et al. [9]. Pichler and Selitsch [17] conclude that the cal-
culation of the parameter values for the α-quantile of the Johnson distribution J(α)
can easily obtained through the following transformation
J(α) = c + d f−1
(N(α)− a
b
),
where N(·) is a standard normal cumulative distribution function.
Other Approaches
As mentioned, in academic literature there are several other methods that suggest
methods of how to calculate the quantile of the profit/loss distribution or how to
determine the distribution exactly:
• Rouvinez [18] uses a trapeziodal rule to invert the characteristic function.
• Britten-Jones and Schaefer [2] suggest an approximation through a χ2 distri-
bution that requires to solve a nonlinear system of equations.
• Cardenas et al. [4] use the Fourier transformation.
• Wilson [21] uses a quadratic programming approach.
5.5 Delta-Normal vs. Delta-Gamma Approaches
As we saw when discussing the delta-normal approach, the approximation error
arising from a “delta-only” representation typically increases with the size of changes
in risk factors. In contrast to the delta-normal approach, the delta-gamma approach
takes the curvature of a non-linear relationship into account by relying on a second-
order Taylor series expansion based on a first and second partial derivatives. The
delta-gamma approach provides a better approximation.
66
This can be illustrated graphically by considering the same numerical example
used earlier when discussing the delta-normal approach. Consider a long position
in a European-style call option. The option’s strike price, the risk-free interest rate,
and the asset’s volatility are given in Table 2.1. Assume the asset trades at $100 two
days before maturity and the call option has one day to maturity. The delta-gamma
approximation and the delta-normal approximation are based on the option’s delta
or delta and gamma two days before maturity. The delta-gamma approximation
was already presented earlier in Figure 4.1. Figure 5.1 compares the delta-normal
approximation with the delta-gamma approximation. The quadratic shape of the
delta-gamma approximation is clearly visible. The result of a quadratic approxima-
tion is that in the case of a long option position, the delta-gamma approximation
leads to values which are higher than the value given by the Black-Scholes model. In
the case of a short position, the delta-gamma approximation leads to values which
are lower than the values given by the Black-Scholes model. However, Figure 5.1 also
shows that the second-order Taylor approximation of the option’s value is closer to
the “true” option value (based on the Black-Scholes option pricing model) than the
delta-normal approximation.
Figure 5.2 compares the absolute approximation errors of the delta-normal
approximation with the delta-gamma approximation. This graph is based on the
same data as the numerical example above. The delta-gamma approximation pro-
vides an approximation that is more accurate in changes in the underlying asset price
over a much larger range than the delta-normal approximation. The approximation
error of a delta-gamma approximation is significantly lower than in the case of the
delta-normal approach. More specifically, the approximation is reliable for smaller
changes in the underlying risk factor. However, there is still an approximation error
when large movements in the underlying risk factor occur.
67
Dowd [6] remarks that although the delta-gamma approximation is more accurate
than the delta-normal method. The delta-gamma method still reveals an approxi-
mation error that might be still too large for some purposes.
However, the increased accuracy in the approximation comes at some cost such
as the loss of normality which is inevitable once we move to second-order approxi-
mations. The following limitations are common to all delta-gamma approaches and
do not depend on the method that is used for calculating the quantile of the distri-
bution.
The improved accuracy by the delta-gamma method comes at the cost of at least
some reduced tractability relative to the delta-normal model. As we have seen, in
using any delta-gamma approach, we might lose normality in our portfolio return
even if changes in the underlying risk factors are normally distributed. The moments
of the profit/loss distribution in the case of the delta-normal and the delta-gamma
approach are summarized in Table 5.1. Table 5.1 shows multiple points. To illustrate
the following points it is necessary that gamma is non-zero, i.e. that the relationship
reveals curvature.
The expected change of the value in the underlying is zero and because of the
linear relationship the expected change in the value of the position in the case of the
delta-normal approach is also zero. Due to the non-linear relationship, the expected
value in the case of the delta-gamma approach is 12S2Γσ2. The sign of the expected
value of the option position will be determined by the relative magnitude and sign
of gamma and whether one is long or short in the option.10
The variance of a change in value in the case of the delta-gamma approach
differs from the variance of the position when using the delta-normal approach by
10As already mentioned, a long position in call or put options always has a positivegamma, whereas a short position in call or put options lead to a negative gamma. Thelack of symmetry has the implication that the VaR now depends on whether one is longor short in the option.
68
Table 5.2: Statistical Properties of the Delta-Normal vs. the Delta-Gamma Approach
Moment Delta-Normal Delta-Gamma Underlying AssetReturn ∆V ∆V ∆xMean 0 1
2S2Γσ2 0
Variance S2δ2σ2 S2δ2σ2 + 12S4Γ2σ4 σ(∆x)
Skewness 0 3S4δ2Γσ4 + S6Γ3σ6 0
an adjustment term 12S4Γ2σ4. If gamma is zero, the position’s variance is the same
as in the delta-normal approach.
Finally, whether the change in value of an option position in the case of using
the delta-gamma approach is positively or negatively skewed depends on whether
the option position is a long or a short position, whereas the distribution in the case
of the delta-normal approach does not reveal asymmetry.
As we saw, non-linear risk exposure leads to a non-normal distribution in posi-
tion returns. This leads to the conclusion that the incidental benefits of normality
which are the ability to translate VaR figures easily from one set of VaR parameters
to another and the ability to infer expected tail losses without any difficulty, are
compromised or lost altogether.
69
90 95 100 105 110Underlying Asset Price (in $)
-5
0
5
10
Val
ue o
f the
Cal
l Opt
ion
(in $
)
"True" Black-Scholes ValueDelta-Gamma ApproximationDelta-Normal Approximation
Figure 5.1: Comparison between the Delta-Normal and the Delta-Gamma Approxi-mation with the Black-Scholes Model.
70
90 95 100 105 110Underlying Asset Price (in $)
0
1
2
3
4
Dev
iatio
n fr
om B
lack
-Sch
oles
val
ue (
in $
)
Delta-Normal Approximation ErrorDelta-Gamma Approximation Error"True" Black-Scholes Deviation
Figure 5.2: Absolute Approximation Error of the Delta-Normal vs. the Delta-GammaApproach.
Chapter 6
Conclusion
VaR models are based on a set of assumptions regarding the distribution of the risk
factors as well as on a set of implicitly made assumptions by standard VaR models.
The changes in risk factors are assumed to follow a multivariate normal distribution.
However, there is empirical evidence that many individual return distributions are
not normal but exhibit fat tails. That means that actual distributions might reveal
a far higher incidence of large market movements than is predicted by a normal dis-
tribution. A normal distribution is more likely to underestimate the risk of extreme
returns.
An implicitly made assumption of standard VaR models is that VaR models rely
on market prices. In particular, these models assume that markets remain liquid
at all times and therefore, assets can be liquidated at prevailing market prices. In
times of market distress liquidity can dry up quickly which can lead to the fact that
positions cannot be sold at all or to dramatically deflated prices. This marking-to-
market approach is questionable if VaR is supposed to represent the worst loss over
a liquidation period.
Another implicit assumption of standard VaR models is that volatilities and
covariances are constant throughout the sample period. The problem is that volatility
and correlation may be highly unstable over the time. In conclusion, VaR estimates
based on normal market conditions may seriously underestimate the risk of a port-
folio when the correlation among risk factors changes. These simplifying assumptions
71
72
can lead to a inaccurate and misleading VaR estimate. Therefore, it is essential to
be aware of the existence of these simplifying assumptions.
Linear approximation models have a number of attractions, such as the fact that
they keep the linearity of the portfolio without adding any new risk factors. The
delta-normal approach provides a tractable way of handling positions with option-
ality that retains the benefits of linear normality. However, the delta-normal method
is only reliable for portfolios with linear risk exposure and can lead to significant
approximation errors for portfolios with non-linear risk exposure. This approxima-
tion error typically increases with the size of changes in risk factors. Delta-normal
approaches are appropriate for linear derivatives and for positions that reveal only
weak curvature and in particular, in cases where the time horizon is very short.
In contrast to the delta-normal approach, the delta-gamma approach takes the
curvature of a non-linear relationship into account as well as provides a significant
improvement over the delta-normal method in the case of non-linear derivatives.
In particular, the delta-gamma approximation is reliable for smaller changes in the
underlying risk factor. However, there is still an approximation error when large
movements in the underlying risk factor occur.
The improved accuracy by the delta-gamma method comes at the cost of at least
some reduced tractability relative to the delta-normal model. Even if changes in the
underlying risk factors are normally distributed, a non-linear relationship leads to
a non-normal probability distribution of the changes in the position’s value. More-
over, this distribution is skewed, meaning that the distribution is asymmetric and
reveals excess kurtosis. This leads to the fact that the incidental benefits of normality
which are the ability to translate VaR figures easily from one set of VaR parameters
to another and the ability to infer expected tail losses without any difficulty, are
compromised or lost.
73
The main difficulty with delta-gamma methods is the estimation of the quantile
of the non-normal profit/loss distribution. There are many different methods dis-
cussed in academic literature for estimating this quantile. The idea discussed in this
paper is a direct approximation of the required quantile based on the Cornish-Fisher
expansion. The normal confidence interval parameter is adjusted for the skewness
and other distortions from normality created by the presence of curvature. How-
ever, suggestions regarding to the number of higher moments that need to be con-
sidered vary. Empirical results by Pichler and Selitsch [17] suggests the use of a
higher moment Cornish-Fisher expansion. In particular, the delta-gamma method,
when using a Cornish-Fisher expansion, provides an approximation which is close to
results calculated by Monte Carlo Simulation methods but is computationally more
efficient.
Glossary of Notation
c Selected confidence level.
Ct Price of a European-style Call option at time t.
d1, d2 Parameters of the Black-Scholes option pricing formulas.
f K×1 column vector containing the current value or level of each risk factor k, fk.
fk Value or level of risk factor k.
Ft Forward or futures price at time t.
FT Delivery price in a forward or futures contract.
K Number of risk factors.
n Number of assets.
N(x) Standard normal cumulative distribution function (CDF) for a variable x. It
defines the cumulative probability that a variable with a standardized normal
variable is less than x.
N ′(x) Probability density function (PDF) for a standard normal variable x.
Pt Price of a European-style Put option at time t.
P Price of a domestic zero-bond with time to maturity τ .
PF Price of a foreign zero-bond with time to maturity τ .
74
75
r Continously compounded domestic risk-free interest rate.
rf Continously compounded risk-free interest rate in a foreign country.
R K × 1 column vector of returns in the risk factors.
St Price of asset underlying a derivative at a general time t.
t A future point in time.
T Time at maturity of a derivative.
V Marked-to-market value of a position.
V (·) Value function that relates risk factors to a value of a derivative.
wi Quantity of asset i.
X Strike or exercise price of a European-style option.
α1−c The 1− c quantile of the assumed distribution.
Γ Gamma of a derivative or portfolio of derivatives.
Γ K ×K Gamma matrix.
δ Delta of a derivative or portfolio of derivatives.
∆St Price change during the period t− 1 to t. Therefore, ∆St = St − St−1.
ηV Kurtosis coefficient of the position.
µ Mean return of the asset.
µV Mean return of the entire position.
ξV Skewness coefficient of the position.
76
ρ Depending on the context, ρ is either the rho of a derivative or describes the
correlation coefficient between two variables.
ρ K ×K correlation matrix.
σ Volatility (i.e. standard deviation) of an asset.
σV Volatility (i.e. standard deviation) of the position.
τ Time horizon τ = T − t.
Θ Theta of a derivative or portfolio of derivatives.
ϑ Vega of a derivative or portfolio of derivatives.
∑K ×K covariance matrix of the risk factor returns.
Bibliography
[1] Black, F.; and Scholes, M. (1973) The Pricing of Options and Corporate Lia-
bilities. Journal of Political Economy, 81, 637-659.
[2] Britten-Jones, M.; and Schaefer S.M. (1999) Non-Linear Value at Risk. Euro-
pean Finance Review 2, 161-167.
[3] Brooks, C.; and Persand, G. (2000) Value at Risk and Market Crashes. Journal
of Risk 2(4).
[4] Cardenas, J.; Fruchard, E.; Koehler, E.; Christophe, M.; and Thomazeau, I.
(1997) VaR: One Step Beyond. Risk, 10 (10), 72-75.
[5] Crouhy M.; Galei D.; and Robert M. (2001) Risk Management. McGraw–Hill.
[6] Dowd, K. (1998) Beyond Value at Risk. John Wiley & Sons.
[7] Duffie, D.; and Pan, J. (1997) An Overview of Value at Risk. Journal of Deriva-
tives, 4(3), 9-49.
[8] Fallon, W. (1996) Calculating Value at Risk. Wharton Financial Institutions
Working Center, Working Paper 96-49.
[9] Hill, I.D.; Hill, R.; and Holder, R.L. (1976) Fitting Johnson Curves by Moments.
Applied Statistics 25 (2), 180-189.
[10] Hull, J.C. (1999) Options, Futures, and other Derivatives. Prentice–Hall.
77
78
[11] Johnson, N.L.; and Kotz, S. (1970) Continous Univariate Distributions. Wiles,
New York.
[12] Jorion, P. (2001) Value at Risk. McGraw-Hill.
[13] J.P.Morgan (1996) RiskMetrics – Technical Document. 4th edition, New York.
[14] J.P.Morgan (1994) Introduction to Risk Metrics. New York.
[15] Markowitz, H. (1952) Portfolio Selection. Journal of Finance 7, 77-91.
[16] Merton, R.C. (1973) Theory of Rational Option Pricing. Bell Journal of Eco-
nomics and Management Science, 4, 141-183.
[17] Pichler, S.; and Selitsch, K. (1999) A Comparison of Analytical VaR Method-
ologies for Portfolios that include Options. Working Paper, Vienna University
of Technology.
[18] Rouvinez, C. (1997) Going Greek with VaR. Risk Magazine, February.
[19] Saunders, A.; and Cornett, M.M. (2003) Financial Institutions Management: A
Risk Management Approach., 4th edition, McGraw-Hill.
[20] Wiener, Z. (1999) Comment on ‘Non-Linear Value-at-Risk’. European Finance
Review 2, 189-193.
[21] Wilson, T.C. (1996) Calculating Risk Capital. The Handbook of Risk Manage-
ment and Analysis. John Wiley & Sons.
[22] Wilson, T.C. (1998) Risk Management and Analysis – Volume 1: Measuring
and Modelling Financial Risk. John Wiley & and Sons.
[23] Zangari, P. (1996) A VaR Methodology for Portfolios that Include Options.
RiskMetrics Monitor, Second Quarter: 4-12.
79
[24] Zangari, P. (1996) How accurate is the Delta-Gamma Methodology? RiskMetrics
Monitor, Third Quarter: 12-29.